fix(agent): per-provider default subagent models; refresh retired Anthropic ids

src/agent/agent_definitions.py

@@ -24,7 +24,10 @@ class AgentDefinition:
     tools: list[str] | None = None  # None or ['*'] means all tools
     source: AgentSource = "built-in"
     base_dir: str = "built-in"
-    model: str | None = None  # None → inherit parent, 'inherit' → force inherit
+    # None → the provider's default subagent model (PROVIDER_INFO
+    # ``subagent_model``; inherits the parent when the provider declares
+    # none). 'inherit' → always the parent's model.
+    model: str | None = None
     permission_mode: PermissionMode | None = None
     max_turns: int | None = None
     background: bool = False
@@ -96,7 +99,8 @@ def _general_purpose_system_prompt(**_kwargs: Any) -> str:
     tools=["*"],
     source="built-in",
     base_dir="built-in",
-    # model intentionally omitted — uses default subagent model
+    # model intentionally omitted — uses the provider's default subagent
+    # model (the cheap fan-out tier; see PROVIDER_INFO subagent_model)
     get_system_prompt=_general_purpose_system_prompt,
 )
 
@@ -160,8 +164,10 @@ def _explore_system_prompt(**_kwargs: Any) -> str:
     omit_clawcodex_md=True,
     # ch08 round-4 (critic M1) — Explore is the fast/cheap read-only agent;
     # TS exploreAgent.ts:77 runs it on Haiku. get_agent_model resolves this
-    # against the session provider and inherits on providers that don't
-    # serve haiku (e.g. DeepSeek), so it is cross-provider safe.
+    # per provider via the ``subagent_tier_models`` tables (anthropic →
+    # claude-haiku-4-5, deepseek → deepseek-v4-flash) and inherits on
+    # providers without a haiku-class mapping, so it is cross-provider
+    # safe.
     model="haiku",
     get_system_prompt=_explore_system_prompt,
 )

src/agent/run_agent.py

@@ -361,11 +361,13 @@ async def run_agent(params: RunAgentParams) -> AsyncGenerator[Message, None]:
 
     # ch08 round-4 WI-1 — per-subagent model resolution (TS getAgentModel,
     # runAgent.ts:340). Resolve the model from the tool param / agent-def /
-    # env, then apply it to a per-subagent provider CLONE. NEVER mutate the
-    # shared session provider: ch07 made Agent concurrency-safe, so N
-    # parallel subagents share params.provider — mutating provider.model
-    # would race across them. copy.copy shares the HTTP client (thread-safe,
-    # per-request model) and gives this subagent its own .model.
+    # env / the provider's default subagent model (PROVIDER_INFO
+    # ``subagent_model``), then apply it to a per-subagent provider CLONE.
+    # NEVER mutate the shared session provider: ch07 made Agent
+    # concurrency-safe, so N parallel subagents share params.provider —
+    # mutating provider.model would race across them. copy.copy shares the
+    # HTTP client (thread-safe, per-request model) and gives this subagent
+    # its own .model.
     turn_provider = params.provider
     try:
         from .agent_model import get_agent_model
@@ -378,6 +380,16 @@ async def run_agent(params: RunAgentParams) -> AsyncGenerator[Message, None]:
         ):
             turn_provider = copy.copy(params.provider)
             turn_provider.model = resolved_model
+            # Debug trace of the routing — nothing in the app configures a
+            # level that shows INFO, so don't pretend otherwise. The
+            # USER-facing surfaces are the Agent tool's result/registry
+            # "model" field and the agent-progress emits (tools/agent.py
+            # resolves the same inputs for reporting).
+            logger.debug(
+                "subagent %s runs on model %s (session model %s)",
+                agent_def.agent_type, resolved_model,
+                getattr(params.provider, "model", None),
+            )
     except Exception:  # noqa: BLE001 — model resolution never blocks a spawn
         logger.debug("subagent model resolution failed; using session model",
                      exc_info=True)

src/command_system/rename_command.py

@@ -93,7 +93,7 @@ async def _generate_session_name(messages: Any) -> str | None:
             default_headers=get_anthropic_custom_headers() or None
         )
         result = client.messages.create(
-            model="claude-3-5-haiku-20241022",
+            model="claude-haiku-4-5",
             max_tokens=100,
             system=_NAME_PROMPT,
             messages=[{"role": "user", "content": text[:4000]}],

src/hooks/exec_agent_hook.py

@@ -59,7 +59,7 @@ async def execute_agent_hook(
         messages = [{"role": "user", "content": user_prompt}]
         system_prompt = AGENT_HOOK_SYSTEM_PROMPT
 
-        effective_model = model or "claude-sonnet-4-20250514"
+        effective_model = model or "claude-sonnet-5"
 
         if hasattr(provider, "chat_async"):
             response = await provider.chat_async(

src/memdir/find_relevant_memories.py

@@ -34,7 +34,7 @@ def _resolve_recall_model(provider: Any) -> str | None:
     TS pins the selector to a small default (``getDefaultSonnetModel``) so a
     turn on an expensive session model doesn't pay full price for the recall
     call. The multi-provider port wires the ``small_fast_model`` setting —
-    BUT its shipped default is an Anthropic id (``claude-3-5-haiku-…``,
+    BUT its shipped default is an Anthropic id (``claude-haiku-4-5``,
     settings/constants.py), which is only valid on the first-party Anthropic
     endpoint. Passing it to a DeepSeek/OpenAI/Minimax session would 400 and
     (since recall swallows errors) silently kill recall every turn (critic

src/models/__init__.py

@@ -27,13 +27,16 @@
     get_context_window_for_model,
     get_model_max_output_tokens,
 )
-from .agent_routing import get_model_for_agent, AgentModelConfig
+# NOTE: the old ``agent_routing`` module (get_model_for_agent /
+# AgentModelConfig) was deleted 2026-08: it was a dead parallel path with
+# inherit-parent semantics that contradicted the shipped resolver
+# (src/agent/agent_model.get_agent_model), and nothing wrote the
+# ``agent_models`` config key it read.
 
 __all__ = [
     "BEDROCK_MODEL_MAP",
     "MODEL_ALIASES",
     "MODEL_CONFIGS",
-    "AgentModelConfig",
     "ModelConfig",
     "canonical_model_name",
     "deprecation_warning",
@@ -42,7 +45,6 @@
     "get_context_window_for_model",
     "get_model_capabilities",
     "get_model_config",
-    "get_model_for_agent",
     "get_model_max_output_tokens",
     "is_model_allowed",
     "resolve_alias",

src/models/aliases.py

@@ -1,31 +1,63 @@
-"""Model alias table matching TypeScript model/aliases.ts."""
+"""Model alias table matching TypeScript model/aliases.ts.
+
+The bare family aliases (``sonnet`` / ``opus`` / ``haiku`` / ``fable``) track
+the CURRENT first-party model of each family — the TS reference resolves
+them through ``getDefaultSonnetModel()``-style functions that are updated at
+every model launch (the ``@[MODEL LAUNCH]`` markers in model.ts). Targets
+below were verified against the live ``GET /v1/models`` catalog 2026-08-12:
+the API no longer serves ``claude-sonnet-4-20250514`` or
+``claude-3-5-haiku-20241022`` (404 ``not_found_error``), so pointing an
+alias at them turns every ``--model sonnet`` / Explore-agent spawn into a
+hard API error.
+
+Note on ``haiku``: the live catalog LISTS only the dated
+``claude-haiku-4-5-20251001``, but the bare ``claude-haiku-4-5`` resolves
+server-side to it (probed live 2026-08-12 — the API's max_tokens 400 names
+the dated id), so the alias uses the bare form like every other current
+target.
+
+Per-provider SUBAGENT tier resolution (e.g. ``haiku`` on a DeepSeek
+session) does not use this table — see ``subagent_model`` /
+``subagent_tier_models`` in ``src/providers/__init__.py`` and
+``src/agent/agent_model.py``.
+"""
 
 from __future__ import annotations
 
 MODEL_ALIASES: dict[str, str] = {
-    # Short names → canonical
-    "sonnet": "claude-sonnet-4-20250514",
-    "opus": "claude-opus-4-20250514",
-    "haiku": "claude-3-5-haiku-20241022",
-
-    # Version aliases
-    "claude-4-sonnet": "claude-sonnet-4-20250514",
-    "claude-4-opus": "claude-opus-4-20250514",
-    "claude-sonnet": "claude-sonnet-4-20250514",
-    "claude-opus": "claude-opus-4-20250514",
-    "claude-haiku": "claude-3-5-haiku-20241022",
-
-    # Legacy aliases
+    # Short names → canonical (current family heads)
+    "sonnet": "claude-sonnet-5",
+    "opus": "claude-opus-5",
+    "haiku": "claude-haiku-4-5",
+    "fable": "claude-fable-5",
+
+    # Version aliases. The explicitly-versioned ``claude-4-*`` spellings
+    # point at the newest LIVE 4.x snapshot of their family (the dated
+    # 2025-05-14 ids they used to target were retired from the API).
+    "claude-4-sonnet": "claude-sonnet-4-6",
+    "claude-4-opus": "claude-opus-4-8",
+    "claude-sonnet": "claude-sonnet-5",
+    "claude-opus": "claude-opus-5",
+    "claude-haiku": "claude-haiku-4-5",
+    "claude-fable": "claude-fable-5",
+
+    # Legacy aliases — explicit historical pins, kept verbatim. These name a
+    # specific retired generation on purpose; the API answers 404 for them,
+    # which is more honest than silently substituting a different model.
     "claude-3.5-sonnet": "claude-3-5-sonnet-20241022",
     "claude-3.5-haiku": "claude-3-5-haiku-20241022",
     "claude-3-sonnet": "claude-3-sonnet-20240229",
     "claude-3-opus": "claude-3-opus-20240229",
     "claude-3-haiku": "claude-3-haiku-20240307",
     "claude-3.7-sonnet": "claude-3-7-sonnet-20250219",
 
-    # Common typos / shortcuts
-    "s4": "claude-sonnet-4-20250514",
-    "o4": "claude-opus-4-20250514",
+    # Common typos / shortcuts. ``s4``/``o4`` mean "sonnet 4"/"opus 4" —
+    # same intent as claude-4-sonnet/claude-4-opus above, so they track the
+    # same newest LIVE 4.x snapshots (their old dated targets were retired
+    # and 404). ``h35`` names a specific retired generation like the legacy
+    # block, and keeps its historical pin.
+    "s4": "claude-sonnet-4-6",
+    "o4": "claude-opus-4-8",
     "h35": "claude-3-5-haiku-20241022",
 }
 

src/models/configs.py

@@ -96,10 +96,9 @@ class ModelConfig:
     # Every 4.x id still matches the earlier ``claude-opus-4`` base first,
     # so their 200K/legacy resolution is unchanged, but two other strings
     # now land here: an unregistered future ``claude-opus-<n>``, AND the
-    # literal ``claude-opus`` — a live MODEL_ALIASES key (aliases.py:15)
-    # that resolves to claude-opus-4-20250514, so ``display_name()`` on the
-    # UNRESOLVED alias now reads "Claude Opus 5". Callers that canonicalize
-    # first (the normal path) are unaffected; test_model_system pins both.
+    # literal ``claude-opus`` — a live MODEL_ALIASES key that (since the
+    # 2026-08 alias refresh) also RESOLVES to claude-opus-5, so the
+    # resolved and unresolved spellings agree on "Claude Opus 5".
     # This inverts the "under-estimate is the safe direction" note above
     # for unknown opus ids, and over-estimating is the worse failure —
     # auto-compact never fires and the request eventually exceeds the real
@@ -122,6 +121,57 @@ class ModelConfig:
         cost_cache_create_per_mtok=6.25,
         cost_cache_read_per_mtok=0.50,
     ),
+    # Claude Sonnet 5 — the current Sonnet tier (the ``sonnet`` alias and
+    # subagent tier target; see ``subagent_tier_models`` in
+    # src/providers/__init__.py). 128K true output cap probed live
+    # 2026-08-12 (``max_tokens: 2000000 > 128000`` from the API's own
+    # 400); the 1M context window is NOT probed — the window probe kept
+    # rate-limiting — and comes from the launch docs plus the Claude-5
+    # family convention (opus-5 / fable-5 / opus-4-8 all ship 1M). $2/$10
+    # per MTok — Sonnet 5's launch pricing was made permanent, so it does
+    # NOT inherit the 4.x sonnet 3/15 tier. Same 32_000 first-attempt wire
+    # ``max_tokens`` convention as the other 1M rows above.
+    #
+    # Placement is load-bearing the same way ``claude-opus-5``'s is: this
+    # row's prefix base is the family-wide ``claude-sonnet``, so it must
+    # sit AFTER ``claude-sonnet-4-20250514`` (the table's first row) —
+    # unknown 4.x sonnet ids (claude-sonnet-4-5-*, -4-6 …) keep resolving
+    # to that conservative 200K row first, and only unregistered future
+    # ``claude-sonnet-<n>`` ids land here.
+    "claude-sonnet-5": ModelConfig(
+        model_id="claude-sonnet-5",
+        display_name="Claude Sonnet 5",
+        context_window=1_000_000,
+        max_output_tokens=32_000,
+        supports_thinking=True,
+        cost_input_per_mtok=2.0,
+        cost_output_per_mtok=10.0,
+        cost_cache_create_per_mtok=2.50,
+        cost_cache_read_per_mtok=0.20,
+    ),
+    # Claude Haiku 4.5 — the anthropic DEFAULT SUBAGENT model and its
+    # ``haiku`` tier target (Explore runs on it, mirroring TS
+    # exploreAgent.ts; cheapest current-gen tier at 1/5 per MTok). The
+    # live catalog LISTS only this dated id, but the bare
+    # ``claude-haiku-4-5`` — the spelling the alias table and subagent
+    # tables use — resolves server-side to it (probed live 2026-08-12).
+    # Window and cap probed the same day: 200K window (``prompt is too
+    # long: … > 200000``) and a 64_000 true output cap (``max_tokens: …
+    # > 64000``), so the 32_000 first-attempt convention leaves the 64K
+    # truncation-escalation exactly at the model's real ceiling. Prefix
+    # base ``claude-haiku-4-5`` also catches the bare spelling and future
+    # dated snapshots.
+    "claude-haiku-4-5-20251001": ModelConfig(
+        model_id="claude-haiku-4-5-20251001",
+        display_name="Claude Haiku 4.5",
+        context_window=200_000,
+        max_output_tokens=32_000,
+        supports_thinking=True,
+        cost_input_per_mtok=1.0,
+        cost_output_per_mtok=5.0,
+        cost_cache_create_per_mtok=1.25,
+        cost_cache_read_per_mtok=0.10,
+    ),
 
     # Claude 3.7 series
     "claude-3-7-sonnet-20250219": ModelConfig(

src/models/model.py

@@ -15,7 +15,7 @@ def resolve_model(name: str) -> str:
     """Resolve a model name/alias to its canonical form.
 
     Steps:
-    1. Resolve alias (e.g. "sonnet" → "claude-sonnet-4-20250514")
+    1. Resolve alias (e.g. "sonnet" → "claude-sonnet-5")
     2. Return canonical name
     """
     canonical = resolve_alias(name)