feat(sdk): central model resolution, GA quality profiles, and HydraFusion opt-in

.changeset/central-model-resolver.md

@@ -0,0 +1,8 @@
+---
+"@bradygaster/squad-sdk": minor
+"@bradygaster/squad-cli": patch
+---
+
+Central model resolver (#1720): `resolveAgentModel` (an alias of `resolveModel`) from `@bradygaster/squad-sdk/agents` adds `config`, `session` and `fallback` sources, applies the exported `MODEL_RESOLUTION_PRECEDENCE` (user-override > config > session > charter > task-auto > default), and returns a `rationale` for every choice (typed as `ModelResolution`; `ResolvedModel.rationale` stays optional so consumer-built values remain valid). `auto` (any case) means "no preference" at every layer, including a per-spawn override. A session-picked model is inherited by every sub-agent spawn with source `session`. When a model listed in `unavailableModels` is in its tier chain, fallback moves only forward to the next available entry; models outside the chain search it from the head; with no candidate left, `omitModel` is set. The returned `fallbackChain` lists only available entries, starting at the selected model when it is in its chain, so retries never move back toward the head. `ModelFallbackExecutor.execute` accepts an optional `{ platformDefault }` callback that runs model-less when the resolution has `omitModel`; it never calls the model callback with an empty id.
+
+The root `resolveModel` export (string-returning, reads `.squad/config.json`) is now a compatible adapter over the central resolver, so it no longer returns a literal `auto`. `AgentLifecycleManager.spawnAgent` now applies persistent `.squad/config.json` preferences (`agentModelOverrides`, `defaultModel`, `economyMode`) and records provenance on `AgentHandle.modelResolution`, OTel span attributes and an `agent:milestone` `model.resolved` event. New helpers: `isModelPreference`, `persistentConfigModel`. Coordinator spawn guidance and `model-selection-reference.md` no longer hard-code model IDs; spawn acknowledgments show model, source and rationale, with catalog IDs kept in a separate appendix. `AgentLifecycleManager.spawnAgent` accepts `sessionModel` and `unavailableModels`; an exhausted chain omits `model` from the session config. Under a cost policy, an explicit over-ceiling pick keeps the forward-only fallback cursor. Availability fallback also honors the active cost ceiling (`config.costPolicy` or `sessionCostPolicy`): catalog-known candidates above it are skipped, so an unavailable out-of-catalog pick (for example `hydrafusion`) falls back only to an in-ceiling entry or omits `model`; an available explicit pick still passes through unchanged.

.changeset/economy-model-fallback.md

@@ -0,0 +1,5 @@
+---
+"@bradygaster/squad-sdk": patch
+---
+
+Allow automatic economy-profile model resolutions to retry available cheaper-tier models after exhausting their selected tier.

.changeset/economy-status-model-ids.md

@@ -0,0 +1,5 @@
+---
+"@bradygaster/squad-cli": patch
+---
+
+Update `squad economy` status examples to match the current economy model substitutions.

.changeset/hydrafusion-opt-in.md

@@ -0,0 +1,5 @@
+---
+"@bradygaster/squad-sdk": minor
+---
+
+Add opt-in HydraFusion support (#2014). `hydrafusion` and `hydrafusion-max` are catalog entries marked `preview` (premium tier) and are excluded from every default chain and GA profile. A new opt-in `hydrafusion` quality profile sends architecture, research, verification, review, security and design tasks to `hydrafusion` first, then the quality-first premium chain; Scribe and mechanical work stay on the fast tier. `hydrafusion-max` is used only when named explicitly. If HydraFusion is unavailable (live model list, or `createSession` rejecting a profile-picked `hydrafusion`, for example under org policy), `AgentLifecycleManager.spawnAgent` falls forward to the premium chain within the cost ceiling instead of failing the spawn, and tells the user once (console warning plus an `agent:milestone` `model.unavailable` event). `ModelRegistry.getRecommendedModels` never recommends preview entries; they stay available through explicit catalog lookup.

.changeset/quality-model-profiles.md

@@ -0,0 +1,9 @@
+---
+"@bradygaster/squad-sdk": minor
+---
+
+Quality-profile and public-catalog aware model selection (#1598). Task-aware selection now runs under a quality profile: `quality-first` (the new default), `balanced` or `economy` (explicit opt-in). Under `quality-first`, architecture, research, verification, review, security, design and visual work use the premium tier; code, prompts, docs, writing and planning use standard; mechanical work (Scribe, git, release) stays fast. The `default` layer now uses the head of the standard chain, or the economy default under `economy`. The profile comes from a session profile (`sessionProfile`), `economyMode: true`, or the persistent `.squad/config.json` `modelProfile` (new `readModelProfile` / `writeModelProfile`, `isModelSelectionProfile`, `MODEL_SELECTION_PROFILES`, `PROFILE_TASK_TIERS`, `resolveModelProfile`). `ResolvedModel.profile`, rationales, the `model.profile` span attribute and the `model.resolved` event report the active profile. Profiles never change explicit user-override, config, session or charter models, and cost ceilings still apply. `economyMode` stays compatible: an explicit `economyMode: false` cancels a persisted `economy` profile. `readEconomyMode` (and `squad economy`) also reports enabled for `modelProfile: "economy"`, and `writeEconomyMode(dir, false)` (`squad economy off`) also removes a persisted `economy` profile, leaving other profiles unchanged.
+
+`resolveAgentModel` and `AgentLifecycleManager.spawnAgent` accept `availableModels`, the runtime's live model list. When supplied, it is authoritative for availability: a missing recommended model falls forward through its chain, and a missing explicit model falls back visibly.
+
+The model catalog, fallback chains (`MODELS.FALLBACK_CHAINS`, `DEFAULT_FALLBACK_CHAINS`, `DEFAULT_CONFIG.models.tiers`) and `ECONOMY_MODEL_MAP` now list only public GA models from the GitHub Copilot supported-models table, with tiers placed by the Copilot model-picker category. Added: `claude-opus-5.5`, `gpt-6.1-sol`, `gpt-6-sol`, `gemini-4-argon`, `claude-sonnet-5.5`, `grok-4.7`, `grok-4.6`, `claude-haiku-5.5`, `gpt-6-luna`, `gemini-3.8-flash`, `mai-code-1.1-flash`. Removed from recommendations, catalog and fallback chains: `claude-opus-4.7`, `claude-opus-4.6`, `claude-sonnet-4.6`, `claude-sonnet-4.5`, `gemini-3.1-pro`. These IDs are still passed through unchanged when a user names one explicitly.

.github/agents/squad.agent.md

@@ -393,11 +393,13 @@ After routing determines WHO handles work, select a **response MODE** (Direct /
 
 ### Per-Agent Model Selection
 
-Resolve a model before every spawn. Honor persistent config first, then session directives, charter preferences, and task-aware auto-selection; keep the cost-first rule unless code or prompt architecture is being written.
+Resolve a model before every spawn with the central resolver precedence (first match wins): **user-override** (explicit per-spawn directive) → **config** (`.squad/config.json` `agentModelOverrides.{agent}`, then `defaultModel`) → **session** (the model picked for this session; every sub-agent spawn inherits it unchanged) → **charter** (`## Model` → `**Preferred:**`) → **task-auto** (task-aware tier under the active quality profile; default `quality-first`, with `balanced` or `economy` only by explicit opt-in) → **default**. `auto` at any layer means no preference; continue to the next. Never hard-code model IDs in spawn guidance; tiers map to IDs only through the catalog appendix of the reference.
 
-Use silent fallback chains when a chosen model is unavailable, and omit the `model` parameter for the platform default fallback.
+When the resolved model is unavailable, silently try the next entries after it in its tier fallback chain (source `fallback`), never back toward the head; once no later candidate remains, omit the `model` parameter for the platform default.
 
-**On-demand reference:** Read `.squad/templates/model-selection-reference.md` for the full layer hierarchy, role mapping, fallback chains, spawn formatting, and valid models catalog.
+Every spawn acknowledgment shows the model, its source and a short rationale: `{emoji} {Name} ({model} · {source}) — {task summary} — {rationale}`.
+
+**On-demand reference:** Read `.squad/templates/model-selection-reference.md` for the full precedence, role-to-tier mapping, fallback rules, spawn formatting, and the model catalog appendix.
 
 ### Per-Agent Reasoning Effort
 
@@ -424,7 +426,7 @@ When the resolved reasoning effort is not `auto` or default, include it in the a
 
 **Spawn output format — show the model choice and effort:**
 
-Follow `.squad/templates/model-selection-reference.md` for the base model-selection rules. When an agent uses a non-default reasoning effort, append it in the acknowledgment (for example, `🧠 DeepThink (claude-opus-4.7-1m-internal · xhigh) — deep architecture analysis`).
+Follow `.squad/templates/model-selection-reference.md` for the base model-selection rules. When an agent uses a non-default reasoning effort, append it in the acknowledgment (for example, `🧠 DeepThink ({model} · charter · xhigh) — deep architecture analysis — charter preference`).
 
 ### Per-Agent Context Tier
 
@@ -451,7 +453,7 @@ When the resolved context tier is not `auto` or default, include it in the agent
 
 **Spawn output format — show the model choice and tier:**
 
-Follow `.squad/templates/model-selection-reference.md` for the base model-selection rules. When an agent uses a non-default context tier, append it in the acknowledgment (for example, `🧠 DeepThink (claude-opus-5 · long context) — 1M-token window for deep architecture analysis`).
+Follow `.squad/templates/model-selection-reference.md` for the base model-selection rules. When an agent uses a non-default context tier, append it in the acknowledgment (for example, `🧠 DeepThink ({model} · charter · long context) — deep architecture analysis — 1M-token window`).
 
 ### Client Compatibility