wizz-method 2.17.0 → 2.17.2

package/package.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://json.schemastore.org/package.json",
   "name": "wizz-method",
-  "version": "2.17.0",
+  "version": "2.17.2",
   "description": "Wizz Method: método de agência orientado por IA em PT-BR",
   "keywords": [
     "agile",

package/src/modules/wizz/_shared/model-ladder.md

@@ -4,7 +4,7 @@ The names `wizz-exec-mechanical`/`wizz-exec-standard`/`wizz-exec-strong`/`wizz-e
 
 ## The recommended session (the boss)
 
-The recommended session model is the Standard tier (Sonnet / gpt-5.6-terra / claude-sonnet-5-5 / gemini-3.5-flash). The boss does common work itself — commands, questions, implementation with a spec — dispatches subagents for isolation or parallel work, arbitrates conflict between executor and reviewer, and takes architecture/product decisions WITH the user.
+The recommended session model is the Standard tier (Sonnet / gpt-6-sol / claude-sonnet-5-5 / gemini-3.5-flash). The boss does common work itself — commands, questions, implementation with a spec — dispatches subagents for isolation or parallel work, arbitrates conflict between executor and reviewer, and takes architecture/product decisions WITH the user.
 
 A session running above Standard (Opus, Fable, Mythos...) absorbs every tier at or below its own model: an Opus session does Strong work itself, and dispatches `wizz-exec-ultra` only after that attempt returns `ESCALAR` per the escalation signals below. A session on the Ultra model (Fable) absorbs Ultra too: it takes strong's `ESCALAR` itself and never dispatches `wizz-exec-ultra`. Review is the exception — it's NEVER absorbed: it always runs on `wizz-exec-review`, pinned to the Strong tier, so in a Fable session review moves off the boss and onto Opus.
 
@@ -14,11 +14,13 @@ How to know who the boss is in Claude Code: the system prompt states the session
 
 | Tier | Subagent | Use | Claude Code | Codex | OpenCode | Gemini CLI |
 |---|---|---|---|---|---|---|
-| Mechanical | `wizz-exec-mechanical` | scanning, rename, repetitive diff, 100% specified change | haiku (maxTurns 50, no effort field) | gpt-5.6-luna, effort low | anthropic/claude-haiku-4-5 | gemini-3.1-flash-lite (max_turns 50, no effort field) |
-| Standard | `wizz-exec-standard` | small/medium defined feature, bug with known cause, tests | sonnet, effort medium (maxTurns 80) | gpt-5.6-terra, effort medium | anthropic/claude-sonnet-5-5 | gemini-3.5-flash (max_turns 80) |
-| Strong | `wizz-exec-strong` | bug with unclear cause after one investigation cycle, refactor spanning modules, schema/auth/payment changes, local technical decision | opus, effort high (maxTurns 120) | gpt-5.6-sol, effort high | anthropic/claude-opus-5-5 | gemini-3.1-pro-preview (max_turns 120) |
+| Mechanical | `wizz-exec-mechanical` | scanning, rename, repetitive diff, 100% specified change | haiku (maxTurns 50, no effort field) | gpt-6-luna, effort low | anthropic/claude-haiku-4-5 | gemini-3.1-flash-lite (max_turns 50, no effort field) |
+| Standard | `wizz-exec-standard` | small/medium defined feature, bug with known cause, tests | sonnet, effort medium (maxTurns 80) | gpt-6-sol, effort medium | anthropic/claude-sonnet-5-5 | gemini-3.5-flash (max_turns 80) |
+| Strong | `wizz-exec-strong` | bug with unclear cause after one investigation cycle, refactor spanning modules, schema/auth/payment changes, local technical decision | opus, effort high (maxTurns 120) | gpt-6-sol, effort high | anthropic/claude-opus-5-5 | gemini-3.1-pro-preview (max_turns 120) |
 | Ultra | `wizz-exec-ultra` | ultra-hard work the Strong tier returned `ESCALAR` on, plus its findings | fable, effort xhigh (maxTurns 150) | gpt-6-astra, effort xhigh | anthropic/claude-fable-5-1 | no ultra tier on this platform |
-| Review | `wizz-exec-review` | review of a diff produced by any tier; pinned to Strong, never inherits the session | opus, effort high (maxTurns 60) | gpt-5.6-sol, effort high | anthropic/claude-opus-5-5 | gemini-3.1-pro-preview (max_turns 60) |
+| Review | `wizz-exec-review` | review of a diff produced by any tier; pinned to Strong, never inherits the session | opus, effort high (maxTurns 60) | gpt-6-sol, effort high | anthropic/claude-opus-5-5 | gemini-3.1-pro-preview (max_turns 60) |
+
+Staying current: Claude Code tiers are aliases that Claude Code points at the newest model, so they never need a bump. Codex has no such alias: on install and on every `wizz update`, the installer reads the machine's Codex models cache (`~/.codex/models_cache.json`) and swaps each pinned Codex model for the newest current model of the same class (fast, workhorse, frontier). The table's Codex column is the fallback, used when the cache is missing, older than the spec, or the newer model lacks the tier's effort. OpenCode and Gemini CLI stay pinned to the table.
 
 OpenCode gets the model only: its per-agent effort field isn't documented, and it has no per-agent turn cap field either. Claude Code and Gemini CLI enforce turn caps natively (`maxTurns` / `max_turns`); Codex and OpenCode have no per-agent turn cap field, so briefs to them should always carry `max_calls`.
 

package/src/modules/wizz/agents/wizz-designer/customize.toml

@@ -41,7 +41,7 @@ principles = [
 [[agent.menu]]
 code = "LP"
 description = "Landing page, hero, conversion page, 3D experience"
-prompt = "Invoke the GLOBAL skill 'wizz-premium-landing-ui-researcher' via the Skill tool to research references/components and build the landing page. Bring the result in easy {communication_language} and show the visual before the code."
+prompt = "Invoke the GLOBAL skill 'wizz-premium-landing-ui-researcher' via the Skill tool to research references/components and build the landing page. Mining runs on live sources opened in this session (load WebSearch/WebFetch first; curl, agent-browser or the local source cache as fallback); with none of them, report BLOCKED instead of recommending from memory. Bring the result in easy {communication_language} and show the visual before the code."
 
 [[agent.menu]]
 code = "UI"
@@ -61,4 +61,4 @@ prompt = "Invoke the GLOBAL skill 'wizz-taste-skill' via the Skill tool to revie
 [[agent.menu]]
 code = "CC"
 description = "Choose components that fit the existing project"
-prompt = "Invoke the GLOBAL skill 'wizz-ui-component-curator' via the Skill tool to analyze the project and recommend compatible components."
+prompt = "Invoke the GLOBAL skill 'wizz-ui-component-curator' via the Skill tool to analyze the project and recommend compatible components. Mining runs on live sources opened in this session (load WebSearch/WebFetch first; curl, agent-browser or the local source cache as fallback); with none of them, report BLOCKED instead of recommending from memory."

package/src/modules/wizz/subagents/codex/wizz-exec-mechanical.toml

@@ -1,9 +1,9 @@
 # Gerado por tools/generate-subagents.mjs a partir de src/modules/wizz/subagents/spec.yaml. Não edite este arquivo direto: edite o spec.yaml e rode `npm run generate:subagents`.
 name = "wizz-exec-mechanical"
 description = """
-Mechanical Wizz Method executor, the cheapest tier. File sweeps, lookups, renames, repetitive diffs, and applying already 100%-specified changes. Dispatch 1 instance when the work fits in 3 files or fewer; parallel instances only when partitioned by file/module (never two on the same file), up to 4 at once. Model on this platform: gpt-5.6-luna, effort low.
+Mechanical Wizz Method executor, the cheapest tier. File sweeps, lookups, renames, repetitive diffs, and applying already 100%-specified changes. Dispatch 1 instance when the work fits in 3 files or fewer; parallel instances only when partitioned by file/module (never two on the same file), up to 4 at once. Model on this platform: gpt-6-luna, effort low.
 """
-model = "gpt-5.6-luna"
+model = "gpt-6-luna"
 model_reasoning_effort = "low"
 developer_instructions = """
 You are a mechanical executor for the Wizz Method. You receive a closed brief and execute it to the letter.

package/src/modules/wizz/subagents/codex/wizz-exec-review.toml

@@ -1,9 +1,9 @@
 # Gerado por tools/generate-subagents.mjs a partir de src/modules/wizz/subagents/spec.yaml. Não edite este arquivo direto: edite o spec.yaml e rode `npm run generate:subagents`.
 name = "wizz-exec-review"
 description = """
-Wizz Method reviewer pinned to the Strong tier (gpt-5.6-sol, effort high), never to the session's model. Receives a diff or an executor's result and reviews correctness, security, regressions, and scope adherence. Doesn't implement: returns a verdict and a list of issues with file:line.
+Wizz Method reviewer pinned to the Strong tier (gpt-6-sol, effort high), never to the session's model. Receives a diff or an executor's result and reviews correctness, security, regressions, and scope adherence. Doesn't implement: returns a verdict and a list of issues with file:line.
 """
-model = "gpt-5.6-sol"
+model = "gpt-6-sol"
 model_reasoning_effort = "high"
 sandbox_mode = "read-only"
 developer_instructions = """

package/src/modules/wizz/subagents/codex/wizz-exec-standard.toml

@@ -1,9 +1,9 @@
 # Gerado por tools/generate-subagents.mjs a partir de src/modules/wizz/subagents/spec.yaml. Não edite este arquivo direto: edite o spec.yaml e rode `npm run generate:subagents`.
 name = "wizz-exec-standard"
 description = """
-Standard Wizz Method executor for day-to-day implementation. Well-defined small/medium features, bug fixes with a known cause, multi-file edits, tests. Same tier as the recommended session model: dispatch it for isolation or parallel work (1 per file/module, up to 4 at once), not to escalate. Model on this platform: gpt-5.6-terra, effort medium.
+Standard Wizz Method executor for day-to-day implementation. Well-defined small/medium features, bug fixes with a known cause, multi-file edits, tests. Same tier as the recommended session model: dispatch it for isolation or parallel work (1 per file/module, up to 4 at once), not to escalate. Model on this platform: gpt-6-sol, effort medium.
 """
-model = "gpt-5.6-terra"
+model = "gpt-6-sol"
 model_reasoning_effort = "medium"
 developer_instructions = """
 You are an implementation executor for the Wizz Method. You receive a brief with a goal and context, and you implement it.

package/src/modules/wizz/subagents/codex/wizz-exec-strong.toml

@@ -1,9 +1,9 @@
 # Gerado por tools/generate-subagents.mjs a partir de src/modules/wizz/subagents/spec.yaml. Não edite este arquivo direto: edite o spec.yaml e rode `npm run generate:subagents`.
 name = "wizz-exec-strong"
 description = """
-Strong Wizz Method executor for hard work. Bugs with no known cause after one investigation cycle, cross-module refactors, schema/auth/payment changes, local technical decisions. The session dispatches it when the work is too hard for the session's own tier. If the work turns out to be ultra hard, it returns ESCALAR with its findings; it never dispatches anything itself. Model on this platform: gpt-5.6-sol, effort high.
+Strong Wizz Method executor for hard work. Bugs with no known cause after one investigation cycle, cross-module refactors, schema/auth/payment changes, local technical decisions. The session dispatches it when the work is too hard for the session's own tier. If the work turns out to be ultra hard, it returns ESCALAR with its findings; it never dispatches anything itself. Model on this platform: gpt-6-sol, effort high.
 """
-model = "gpt-5.6-sol"
+model = "gpt-6-sol"
 model_reasoning_effort = "high"
 developer_instructions = """
 You are a senior executor for the Wizz Method. You receive a brief with a goal and context for hard work, where the root cause or the path forward isn't clear yet.

package/src/modules/wizz/subagents/spec.yaml

@@ -169,12 +169,15 @@ models:
     strong: anthropic/claude-opus-5-5
     ultra: anthropic/claude-fable-5-1
     review: anthropic/claude-opus-5-5
+  # GPT-6 has three models for five tiers (2026-10-05, Codex models cache):
+  # luna = fast/cheap, sol = workhorse, astra = frontier. Strong/review
+  # reuse sol at higher effort; ultra stays on astra.
   codex:
-    mechanical: gpt-5.6-luna
-    standard: gpt-5.6-terra
-    strong: gpt-5.6-sol
+    mechanical: gpt-6-luna
+    standard: gpt-6-sol
+    strong: gpt-6-sol
     ultra: gpt-6-astra
-    review: gpt-5.6-sol
+    review: gpt-6-sol
 
 # Codex `model_reasoning_effort`.
 codex_effort:

package/src/skills-lib/wizz-premium-landing-ui-researcher/SKILL.md

@@ -16,6 +16,10 @@ Check for `{output_folder}/design-state.yaml` (default `_wizz-output/`). Cheap c
 - **It does not exist, and the work is non-trivial:** invoke `skill:wizz-design-needs` first. Mining without a map is how a project ends up depending on the user to name every effect, one request at a time.
 - **It does not exist, and the work is one isolated component the user named:** proceed, and say the map was not created.
 
+## Gate 0.5: open the access (always, before the first search)
+
+Read [web-access-and-evidence](references/web-access-and-evidence.md) and run its Step 0: load the web tools (on Claude Code, WebSearch and WebFetch are deferred and need `ToolSearch`), then fall back to `curl`, `agent-browser` and the local source cache. No web and no cache means BLOCKED, never research from memory. Before delivering, every recommended component passes that file's evidence gate; the rest are UNVERIFIED and are never recommended.
+
 ## Gate 1: Source-First (always, before any UI code)
 
 This skill exists because writing shaders, animations, hovers, and components from scratch is NOT the way. The way is to curate mature components, animations, and shaders from professional sources (the user's `modelos lp/`, React Bits, Cult UI, Ali Imam, Watermelon, StyleUI, Skiper UI, Componentry, and public registries) and adapt them to the brand.
@@ -64,5 +68,6 @@ Classify the project into one of 5 levels (rules, allowed/forbidden motion, and
 | Base prompts (complete landing page and 3D scroll hero) | [prompt-templates](references/prompt-templates.md) |
 | Final response format, Quality Bar, and self-check before closing the delivery | [output-format-and-quality](references/output-format-and-quality.md) |
 | Direct links to every authorized source | [source-links](references/source-links.md) |
+| Opening web access before mining, BLOCKED, and the evidence gate before delivering | [web-access-and-evidence](references/web-access-and-evidence.md) |
 
 Final rule: never put real keys inside the skill (environment variables only), and never deliver a landing page that looks like a template with no positioning.

package/src/skills-lib/wizz-premium-landing-ui-researcher/references/component-sources.md

@@ -11,7 +11,7 @@ Research flow (public sources first; log evidence per [source-first-protocol](so
 3. public shadcn registries;
 4. taste/motion sources (Impeccable, Taste Skill, Design Motion Principles, MotionSites, Vibe Code Components, Refero Styles with a free DESIGN.md);
 5. open visual sources (Landing Love, Godly, Design Spells, Mobbin, Refero, ScreensDesign, DesignVault, Spline, Unicorn Studio);
-6. coherent strategic hypotheses, when there is no external access.
+6. with no web and no local cache: BLOCKED, per [web-access-and-evidence](web-access-and-evidence.md). Strategic hypotheses can shape the direction, never stand in for a component source.
 
 ## Mining Evidence
 

package/src/skills-lib/wizz-premium-landing-ui-researcher/references/source-first-protocol.md

@@ -56,7 +56,7 @@ Log in the existing design artifact (or create `design-research.md` in the proje
 |---|---|---|---|---|---|---|
 | Item actually inspected | Terms used and catalog | Real URLs or paths | SHA/tag; date if there's no version | Verified or pending | Tested / static / unavailable | Accepted/rejected and reason |
 
-Also include searches with no results and unavailable sources. Never fill in component names, paths, license, or test results by assumption.
+Also include searches with no results and unavailable sources. Never fill in component names, paths, license, or test results by assumption. A row missing any of the five fields in [web-access-and-evidence](web-access-and-evidence.md) is UNVERIFIED and cannot be the recommendation.
 
 Show 1-3 compatible candidates per component, with a recommendation and adaptation cost. If the user has already authorized the implementation and defined the direction, choose the best option within that scope and proceed. Only ask when a material decision about direction, paid access, or scope change is missing.
 

package/src/skills-lib/wizz-premium-landing-ui-researcher/references/web-access-and-evidence.md

@@ -0,0 +1,40 @@
+# Web access and evidence gate
+
+Shared by `wizz-premium-landing-ui-researcher` and `wizz-ui-component-curator`. Read it before the first search. It covers two moments: opening access before mining, and checking evidence before delivering.
+
+Mining from memory is the failure this file exists to prevent. Component names, URLs, licenses and APIs the model "remembers" are hypotheses, and a hypothesis delivered as research is worse than no research: the user builds on it without knowing.
+
+## Step 0: open the access (before any query)
+
+A tool you do not see is not a tool you do not have. In a subagent, the web tools often start unloaded.
+
+1. **Load the web tools.** On Claude Code, WebSearch and WebFetch are deferred: load them with `ToolSearch` (`select:WebSearch,WebFetch`) before deciding they are missing. Do the same for an Exa or scrapling MCP if the session lists one. On other platforms, use that platform's native web search and fetch.
+2. **Fallbacks, in order:** `curl -sL <url>` through the shell (raw HTML, registry JSON, `raw.githubusercontent.com` source files), then `agent-browser` (also the tool for watching hover and motion live).
+3. **Local source cache:** `~/.claude/design-sources/` holds real repositories (React Bits, Cult UI, Ali Imam, Watermelon, StyleUI). Reading a file there counts as evidence, as long as you record the path and the revision (`git -C <repo> rev-parse --short HEAD`) and say the cache may be behind the remote.
+4. **Record the outcome** in one line before mining: which access worked (`web`, `curl`, `agent-browser`, `cache only`) and which failed, with the error.
+
+If none of the four works, stop and return:
+
+```text
+BLOCKED: no web access and no local source cache.
+Tried: <tool → error, one per line>
+Unblock: <what the user can do, e.g. enable WebFetch, install agent-browser, clone the sources into ~/.claude/design-sources/>
+```
+
+BLOCKED is a complete, correct answer. Do not deliver component recommendations instead of it, not even marked "(verify)". Direction work that needs no source (site level, motion budget, style read of the project) can still be delivered next to BLOCKED, labelled as such.
+
+## Evidence gate (before delivering)
+
+Every candidate needs all five, collected in this session:
+
+| Field | Counts | Does not count |
+|---|---|---|
+| URL or path | The exact demo and the exact source file or registry item you opened | A catalog homepage, a search snippet, a URL you did not open |
+| Date | Today's date or the repo revision (SHA/tag) | "recent", "current" |
+| License | Read in the repo or registry | Assumed from the catalog's reputation |
+| Dependencies | Read from the actual imports or registry JSON | Guessed from the component's name |
+| Source read | The code file was opened and its imports checked | The demo alone |
+
+A candidate missing any field is **UNVERIFIED**. UNVERIFIED items go in a separate section of the report, with what is missing and why. They are never recommended and never placed in the top 1-3, and the implementation plan never depends on them. If every candidate for a need is UNVERIFIED, the need is reported as `unmet`, not filled.
+
+Before closing, check the delivery against this gate once: every recommended item has its five fields in the research log (`design-research.md` or the project's design artifact), and the access line from Step 0 is in the report.

package/src/skills-lib/wizz-ui-component-curator/SKILL.md

@@ -21,7 +21,7 @@ If `{output_folder}/design-state.yaml` exists, the need being worked is already
 
 Not there? Look before deciding it does not exist: `find . -name design-state.yaml -not -path '*/node_modules/*'`. A map found anywhere else is still the map: use it, and tell the user it is off-place (it belongs in `{output_folder}`).
 
-With no state file, continue normally.
+With no state file and more than one need in the request (a polish pass, "effects and animations", a dashboard), run `wizz-design-needs` first and mine against the map it writes. With no state file and one isolated component the user named, continue, and say the map was not created.
 
 ## 1. Read and infer the style DNA
 
@@ -42,7 +42,7 @@ Do not ask the user to describe the style unless the project has no inspectable
 
 Start with existing project components. Then search the public sources. The full authorized list (React Bits, Cult UI, Ali Imam, Watermelon, StyleUI, Skiper UI, Bklit UI, Componentry, shadcn/ui, plus the reference galleries) lives in one place: [component-sources](../wizz-premium-landing-ui-researcher/references/component-sources.md). Read it instead of relying on memory. Compare at least two suitable sources per need.
 
-Use whatever search tools the session has: web search (Exa when configured, WebSearch otherwise), WebFetch, scrapling, and `agent-browser` to watch hover and motion live. A tool that errors or hangs is not the end of the research: record the failure, move to the next tool, and never stop without reporting what was and was not checked.
+Before the first query, open the access as described in [web-access-and-evidence](../wizz-premium-landing-ui-researcher/references/web-access-and-evidence.md): load the web tools (on Claude Code, WebSearch and WebFetch are deferred and need `ToolSearch`), then fall back to `curl`, `agent-browser` and the local source cache. Search with whatever that step opened: web search (Exa when configured, WebSearch otherwise), WebFetch, scrapling, and `agent-browser` to watch hover and motion live. A tool that errors or hangs is not the end of the research: record the failure, move to the next tool, and never stop without reporting what was and was not checked. With no web and no cache, return BLOCKED instead of recommending from memory.
 
 Open the exact demo and source file/registry item. Check the actual imports, license, dependencies, stack variant and revision; use current official documentation (Context7 when available) for API compatibility. Inspect animation/interaction in the available browser; label static-only inspection honestly. If a source fails, record the failure and continue elsewhere.
 
@@ -57,6 +57,8 @@ Use the inferred style DNA as the search filter. Prefer:
 
 Do not present unranked lists. Shortlist internally, then surface only the strongest matches.
 
+Only candidates that pass the evidence gate in [web-access-and-evidence](../wizz-premium-landing-ui-researcher/references/web-access-and-evidence.md) can be shown here. The rest are UNVERIFIED and go to the evidence section.
+
 For each candidate, say briefly:
 - why it fits
 - what role it plays
@@ -131,4 +133,4 @@ Up to 2 backup options only if genuinely different and useful.
 What changes, which files, any dependency or adaptation work.
 
 ## Research evidence
-Exact candidate demo/source links, revision/date, license and dependencies, visual inspection status, decisions and unavailable sources. State what still needs a user decision.
+The access line (what worked, what failed), then exact candidate demo/source links, revision/date, license and dependencies, visual inspection status, decisions and unavailable sources. UNVERIFIED items listed apart, with what is missing. State what still needs a user decision.

package/tools/installer/core/project-instructions.js

@@ -22,8 +22,9 @@ const INSTRUCTION_FILES = ['CLAUDE.md', 'AGENTS.md'];
 // claude-haiku-4-5 continua atual (não há Haiku mais novo em 2026-09-30).
 // `claude-(opus|sonnet)-5` sem sufixo foi trocado por 5-5; o lookahead evita
 // casar o próprio `-5-5`.
+// A família GPT 5.x (Codex) saiu da escada em 2026-10-05: a família GPT-6 assumiu.
 const RETIRED_MODEL_RE =
-  /\bclaude-(?:instant|2|3|3-5|3-7)(?:-[a-z0-9]+)*\b|\bclaude-(?:opus|sonnet)-4(?:-\d+)?\b|\bclaude-(?:opus|sonnet)-5\b(?!-)/gi;
+  /\bclaude-(?:instant|2|3|3-5|3-7)(?:-[a-z0-9]+)*\b|\bclaude-(?:opus|sonnet)-4(?:-\d+)?\b|\bclaude-(?:opus|sonnet)-5\b(?!-)|\bgpt-5(?:\.\d+)?(?:-[a-z]+)*\b/gi;
 
 const LEGACY_RULES = [
   { id: 'gsd', re: /\bGSD\b|get-shit-done|\/gsd[:-]/, why: 'GSD foi substituído pelas fases e checkpoints do Wizz Method' },

package/tools/installer/core/subagent-models.js

@@ -0,0 +1,112 @@
+// Modelo dos subagentes resolvido na instalação (2026-10-05).
+//
+// O spec.yaml pina um modelo por tier e plataforma. No Claude Code isso é um
+// apelido (haiku/sonnet/opus/fable) que o próprio Claude Code aponta para a
+// versão mais nova, então o arquivo passa intacto. O Codex não tem apelido de
+// "mais novo": aqui o installer lê o cache de modelos do Codex desta máquina
+// (~/.codex/models_cache.json) e troca o modelo pinado pelo mais novo da mesma
+// classe (barato / do dia a dia / fronteira). Qualquer dúvida mantém o pinado:
+// sem cache, cache mais velho que o spec (o pinado nem aparece nele), esforço
+// que o modelo novo não suporta, ou JSON em formato inesperado.
+//
+// OpenCode e Gemini seguem pinados: falta uma fonte local confiável da lista
+// de modelos deles.
+
+const fs = require('node:fs');
+const os = require('node:os');
+const path = require('node:path');
+
+const CLASS_PATTERNS = {
+  fast: /\b(fast|affordable|efficient|lightweight|cheap|mini|nano)\b/i,
+  frontier: /\b(frontier|most capable|most demanding|most intelligent)\b/i,
+};
+const OUTDATED = /\b(older|legacy|deprecated)\b/i;
+
+function defaultCodexCachePath() {
+  return process.env.CODEX_MODELS_CACHE || path.join(os.homedir(), '.codex', 'models_cache.json');
+}
+
+function loadCodexCache(cachePath = defaultCodexCachePath()) {
+  try {
+    return JSON.parse(fs.readFileSync(cachePath, 'utf8'));
+  } catch {
+    return null;
+  }
+}
+
+// Classe pela descrição; um modelo antigo ("Older fast and efficient model")
+// ainda casa a classe certa porque o padrão ignora o "Older".
+function classOf(entry) {
+  const text = `${entry.slug} ${entry.description || ''}`;
+  if (CLASS_PATTERNS.frontier.test(text)) return 'frontier';
+  if (CLASS_PATTERNS.fast.test(text)) return 'fast';
+  return 'workhorse';
+}
+
+function isCurrent(entry) {
+  return entry.visibility === 'list' && !entry.upgrade && !OUTDATED.test(entry.description || '');
+}
+
+function supportsEffort(entry, effort) {
+  if (!effort) return true;
+  const levels = Array.isArray(entry.supported_reasoning_levels) ? entry.supported_reasoning_levels : [];
+  return levels.some((level) => level && level.effort === effort);
+}
+
+/**
+ * Modelo Codex mais novo da mesma classe do pinado, ou o próprio pinado.
+ * @param {string} pinned - modelo do spec
+ * @param {string|undefined} effort - model_reasoning_effort do arquivo
+ * @param {Object|null} cache - conteúdo de models_cache.json
+ */
+function resolveCodexModel(pinned, effort, cache) {
+  const models = cache && Array.isArray(cache.models) ? cache.models.filter((m) => m && typeof m.slug === 'string') : null;
+  if (!models) return pinned;
+  const pinnedEntry = models.find((m) => m.slug === pinned);
+  // O pinado fora do cache = cache mais velho que o spec: nunca rebaixar.
+  if (!pinnedEntry) return pinned;
+  const wanted = classOf(pinnedEntry);
+  const newest = models
+    .filter((m) => isCurrent(m) && classOf(m) === wanted)
+    .sort((a, b) => (a.priority ?? Infinity) - (b.priority ?? Infinity))[0];
+  if (!newest || !supportsEffort(newest, effort)) return pinned;
+  return newest.slug;
+}
+
+function readModel(content) {
+  const match = content.match(/^model = "([^"]+)"$/m) || content.match(/^model: (\S+)$/m);
+  return match ? match[1] : null;
+}
+
+function replaceModel(content, from, to) {
+  return from && to && from !== to ? content.split(from).join(to) : content;
+}
+
+/**
+ * Conteúdo do subagente com o modelo resolvido para esta máquina.
+ * @param {string} content - arquivo gerado a partir do spec
+ * @param {string|undefined} platform - agents_source_subdir (codex, gemini...)
+ * @param {Object} [opts] - codexCache (objeto) ou codexCachePath
+ */
+function renderSubagent(content, platform, opts = {}) {
+  if (platform !== 'codex') return content;
+  const pinned = readModel(content);
+  if (!pinned) return content;
+  const effort = (content.match(/^model_reasoning_effort = "([^"]+)"$/m) || [])[1];
+  const cache = opts.codexCache === undefined ? loadCodexCache(opts.codexCachePath) : opts.codexCache;
+  return replaceModel(content, pinned, resolveCodexModel(pinned, effort, cache));
+}
+
+/**
+ * Dois arquivos iguais a menos do modelo: um arquivo instalado com outro
+ * modelo resolvido continua sendo do Wizz (não precisa de backup).
+ */
+function sameExceptModel(a, b) {
+  if (a === b) return true;
+  const modelA = readModel(a);
+  const modelB = readModel(b);
+  if (!modelA || !modelB) return false;
+  return a.split(modelA).join('<wizz-model>') === b.split(modelB).join('<wizz-model>');
+}
+
+module.exports = { renderSubagent, resolveCodexModel, sameExceptModel, loadCodexCache };

package/tools/installer/ide/_config-driven.js

@@ -6,6 +6,7 @@ const prompts = require('../prompts');
 const { loadCsvCached } = require('../core/yaml-cache');
 const { WIZZ_FOLDER_NAME } = require('./shared/path-utils');
 const { findExistingWizzDir } = require('../core/wizz-dir');
+const { renderSubagent, sameExceptModel } = require('../core/subagent-models');
 const { getInstalledCanonicalIds, isWizzOwnedEntry } = require('./shared/installed-skills');
 
 // Reserved OpenCode slash commands. A skill whose canonicalId collides with
@@ -567,7 +568,8 @@ class ConfigDrivenIdeSetup {
           }
           const existing = await fs.readFile(targetFile, 'utf8');
           const source = await fs.readFile(sourcePath, 'utf8');
-          if (existing !== source) {
+          // Same file with another resolved model is still Wizz-generated.
+          if (!sameExceptModel(existing, source)) {
             const backupPath = await this._backupSubagentFile(targetFile, options, report);
             if (!backupPath) {
               report.preserved++;
@@ -576,7 +578,10 @@ class ConfigDrivenIdeSetup {
             report.backups++;
           }
         }
-        await fs.copy(sourcePath, targetFile, { overwrite: true });
+        const rendered = renderSubagent(await fs.readFile(sourcePath, 'utf8'), config.agents_source_subdir, {
+          codexCachePath: options.codexModelsCachePath,
+        });
+        await fs.writeFile(targetFile, rendered);
         count++;
         report.installed++;
       } catch (error) {
@@ -930,7 +935,7 @@ class ConfigDrivenIdeSetup {
       // Exact source equality is the only proof that a source-backed file is
       // unchanged. Without a matching source, both current and legacy tiers
       // may contain customizations and must be backed up before removal.
-      const generated = sourceContent !== undefined && content === sourceContent;
+      const generated = sourceContent !== undefined && sameExceptModel(content, sourceContent);
       if (!generated) {
         const backupPath = await this._backupSubagentFile(filePath, options, report);
         if (!backupPath) {