ksh999000/Claude-Code-Game-Studios
1
0
Fork 0
mirror of https://github.com/Donchitos/Claude-Code-Game-Studios.git synced 2026-10-16 01:37:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

Release 1.1.2 — the default path works end to end

Browse Source 
Fix release for 1.1.1: the minimal path end to end, engine commands that
run on Godot, Unity and Unreal, hook warnings you can see, and verdicts
that say when they could not check. Details in CHANGELOG.md; upgrade
steps in UPGRADING.md (v1.1.1 → v1.1.2).

Fixes #67
Fixes #131

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WuJn2FdynEybfL5KC697nJ
This commit is contained in:
Donchitos
2026-09-29 21:32:19 +10:00
co-authored by Claude Opus 5.5
parent 7ed2c3e9c4
commit b21fa0f7f2
383 changed files with 22888 additions and 10618 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.claude/docs/model-tiers.md
+78 -25
View File
@@ -1,32 +1,47 @@
# Model Tier Assignment
> ### DECLARED, NOT CURRENTLY APPLIED — skills only
> ### When a skill's `model:` is used — skills only
>
> **Claude Code reads `model:` from a skill's frontmatter and then serves the
> skill from the session's model anyway.** This has been measured directly
> against session transcripts on several Claude Code versions, and held every
> time. To re-test it on a newer version, run a skill that declares a tier
> different from the session model and check which model the transcript
> attributes the skill's turns to.
> A skill's `model:` overrides the session model for the rest of the turn the
> skill runs in, and your next prompt returns to the session model. Whether the
> override happens depends on how the skill starts (measured on Claude Code
> 2.1.282):
>
> So the skill tiers below are **intent, not behaviour**. Setting one changes
> nothing today. They are kept because they are correct descriptions of what each
> skill needs, and because they cost nothing to carry if Claude Code honours them
> later — but do not budget on them, and never tell a user a tier is saving them
> money.
> | How the skill starts | Is `model:` used? |
> |---|---|
> | You type `/skill-name` in Manual (`default`) mode | Yes |
> | You type `/skill-name` in auto mode | `sonnet` and `opus`: yes. `haiku`: no — auto mode does not support Haiku, so the session keeps its model (documented) |
> | Claude starts the skill itself, through the Skill tool | No — the session model runs it (measured; the docs do not say) |
>
> Two things follow:
>
> - **An `opus` skill costs more in a Sonnet session.** Typing
> `/architecture-review`, `/gate-check` or `/review-all-gdds` moves that turn
> to Opus.
> - **A `haiku` skill saves nothing in auto mode** (this project's settings start
> terminal sessions in Manual, but a user can switch) — or when Claude starts
> it. Never tell a user a `haiku` tier is saving them money.
>
> Because the override ends at your next prompt, a skill that stops to ask you
> something in chat finishes on the session model after you answer.
>
> **This does NOT apply to agents.** `model:` in `.claude/agents/*.md` is a
> different mechanism, spawned as a separate session rather than executed inline.
> It has not been measured here, so treat the agent section below as unverified
> in either direction rather than assuming it behaves the same way.
> different mechanism, spawned as a separate session rather than executed inline,
> and Claude Code documents it as applied: it is second in the subagent model
> order, after a `model` Claude passes for one invocation. See the agent section
> below.
Skills and agents are assigned tiers by task complexity:
Skills and agents are assigned tiers by task complexity. A tier is written as
the alias `model:` takes, never as a model ID: Claude Code resolves each alias
to the current model in that family, so this table does not change when a new
model ships.
| Tier | Model | When to use |
|------|-------|-------------|
| **Haiku** | `claude-haiku-4-5-20251001` | Read-only status checks, formatting, simple lookups — no creative judgment needed |
| **Sonnet** | `claude-sonnet-5` | Implementation, design authoring, analysis of individual systems — default for most work |
| **Opus** | `claude-opus-5` | Multi-document synthesis, high-stakes phase gate verdicts, cross-system holistic review |
| Tier | `model:` | When to use |
|------|----------|-------------|
| **Haiku** | `haiku` | Read-only status checks, formatting, simple lookups — no creative judgment needed |
| **Sonnet** | `sonnet` | Implementation, design authoring, analysis of individual systems — default for most work |
| **Opus** | `opus` | Multi-document synthesis, high-stakes phase gate verdicts, cross-system holistic review |
| **Session** | `inherit` | Agents only — runs on whatever model the session runs |
Skills with `model: haiku` (5): `/help`, `/onboard`, `/project-stage-detect`,
`/scope-check`, `/sprint-status`
@@ -49,7 +64,45 @@ this repo declares a tier, and the lists above are kept in step with what the
`SKILL.md` files declare. That is a check on two descriptions agreeing; it
proves nothing about which model actually runs.
**Agent model tiers.** Orchestrator-spawned agents (`team-*`, `/review-all-gdds`)
use `model: inherit` so a fixed smaller model can't overflow when the parent runs
a large-context tier. Directors stay `opus`, `community-manager`/`devops-engineer`
stay `haiku`, and non-orchestrator agents stay pinned at `sonnet` as a floor.
**Agent model tiers.** 30 agents use `model: inherit` and run on the session's
model. 19 are pinned:
- **opus (3):** `creative-director`, `producer`, `technical-director`. Under an
Opus session an `opus` pin runs on the session's exact model, `[1m]` included,
so it never gets a smaller window than its parent.
- **sonnet (16):** `lead-programmer`, `prototyper`, `devops-engineer`,
`community-manager` and the 12 engine sub-specialists:
`godot-csharp-specialist`, `godot-gdextension-specialist`,
`godot-gdscript-specialist`, `godot-shader-specialist`,
`ue-blueprint-specialist`, `ue-gas-specialist`, `ue-replication-specialist`,
`ue-umg-specialist`, `unity-addressables-specialist`,
`unity-dots-specialist`, `unity-shader-specialist`, `unity-ui-specialist`.
`/dev-story` and `/code-review` spawn the sub-specialists named in
`project.yaml`'s `specialists` block (which `/setup-engine` writes); `/team-ui` spawns
`unity-ui-specialist` and `ue-umg-specialist`; `/team-release` spawns
`devops-engineer` and `community-manager`, and `/team-live-ops` spawns
`community-manager`.
- **haiku (0):** no agent. The Haiku row above is for read-only lookups and
formatting; every agent drafts, writes or judges. Building CI pipelines and
writing player-facing text are Sonnet work, as they are for `/patch-notes`.
Every agent not named above is `inherit`. The `model:` line in the agent's own
file is what runs, and this list is kept in step with those files.
A pin gets that model's context window, whatever the session runs. On the
Anthropic API `sonnet` currently resolves to Sonnet 5, which always has a 1M
window, so a `sonnet` pin is not smaller than a 1M session there; `haiku`
resolves to Haiku 4.5 and its 200K window, and behind an LLM gateway a `sonnet`
pin can be smaller too. A subagent starts from
the brief it is given, not the parent's conversation, so this only matters when
a brief is very large. If a pinned agent does overflow, run every subagent on
the session's model with one line in your `.claude/settings.local.json`
(Claude Code v2.1.257 or later):
```json
{ "env": { "CLAUDE_CODE_SUBAGENT_MODEL_FORCE": "1" } }
```
It applies to **every** pin, the three `opus` directors included: under a
Sonnet session they run on Sonnet too. `CLAUDE_CODE_SUBAGENT_MODEL` alone does
not do it: it applies only to agents that set no model of their own.