feat(models): add claude-opus-5 + fix interactive /effort on Anthropic

eval/harbor/README.md

@@ -54,7 +54,7 @@ PYTHONPATH=$PWD/eval/harbor harbor run \
 NOTE: hub datasets namespace task names — filters must match the full
 name: `-i 'terminal-bench/fix-git'` (or use a glob: `-i '*fix-git*'`).
 
-## Evaluate with claude-opus-4-8 on a Claude subscription
+## Evaluate with claude-opus-5 on a Claude subscription
 
 Uses your Claude Pro/Max subscription (OAuth) instead of an API key.
 One-time prerequisite on the host: `clawcodex login` (writes
@@ -70,9 +70,10 @@ covers only the main loop).
 PYTHONPATH=$PWD/eval/harbor harbor run \
   --dataset terminal-bench/terminal-bench-2-1 \
   --agent clawcodex_agent:Clawcodex \
-  --model anthropic/claude-opus-4-8 \
+  --model anthropic/claude-opus-5 \
   --ak subscription=true \
   --ak effort=high \
+  --ak source=git+https://github.com/agentforce314/clawcodex@main \
   --jobs-dir eval/harbor/jobs \
   --n-concurrent 2
 ```
@@ -82,10 +83,21 @@ Notes:
   resources. Keep Harbor's default timeout policy for leaderboard-comparable
   runs; use timeout multipliers only for explicitly labeled diagnostics.
 - `effort=high` maps to `clawcodex --effort high` →
-  `output_config.effort` on effort-capable models (Opus 4.6/4.8,
+  `output_config.effort` on effort-capable models (Opus 5, Opus 4.6/4.8,
   Sonnet 4.6, Fable 5). Requires clawcodex > 1.2.1 in the container —
-  until the next PyPI release, add
-  `--ak source=git+https://github.com/agentforce314/clawcodex@main`.
+  until the next PyPI release, keep the
+  `--ak source=git+https://github.com/agentforce314/clawcodex@main` above.
+- **claude-opus-5 needs a clawcodex that registers it** (the model tables
+  in `src/models/configs.py`, `src/services/pricing.py`, and the three
+  gates in `src/query/query.py`). On a build without it, opus-5 falls off
+  the adaptive-thinking allowlist and the request carries
+  `thinking={"type": "enabled", "budget_tokens": …}` — removed on Opus 5,
+  so **every request 400s** — while `--effort` is silently dropped and the
+  context window is assumed to be 200K instead of 1M (early compaction,
+  the handicap #730 removed for opus-4-8). PyPI 1.2.1 predates this, so
+  `source=` is required; verify the branch/SHA you point it at actually
+  carries the registration before starting a full run, and prefer a
+  commit SHA over `@main` for anything you plan to compare later.
 - Subscription rate limits are shared with your interactive Claude
   usage — keep `--n-concurrent` low (2-4) and consider
   `--max-retries 2 --retry-include ApiRateLimitError`.
@@ -105,7 +117,7 @@ token), so multi-hour jobs work without a manually exported token:
 PYTHONPATH=$PWD/eval/harbor harbor run \
   --dataset terminal-bench/terminal-bench-2-1 \
   --agent claude_code_subscription:ClaudeCodeSubscription \
-  --model anthropic/claude-opus-4-8 \
+  --model anthropic/claude-opus-5 \
   --ak reasoning_effort=high \
   --jobs-dir eval/harbor/jobs \
   --n-concurrent 2
@@ -114,8 +126,16 @@ PYTHONPATH=$PWD/eval/harbor harbor run \
 Notes:
 - Effort uses the parent agent's kwarg name: `--ak reasoning_effort=`
   (low|medium|high|xhigh|max).
+- The model is forwarded as the `ANTHROPIC_MODEL` env var with the
+  provider prefix stripped (`harbor/agents/installed/claude_code.py`
+  :1372-1386 — so don't expect a `--model` on the exec line when
+  debugging), and effort as the `--effort` CLI flag. Neither is gated on
+  a model table anywhere in the adapter, and the bootstrap installs the
+  latest official CLI in every container, so moving this arm to a new
+  model is the model string alone — no adapter or clawcodex change.
 - Pin the CLI to a leaderboard row's version with `--ak version=2.1.205`
-  (default: latest).
+  (default: latest). A pin predating a model's support will fail on that
+  model, so re-check the pin when changing `--model`.
 - `CLAUDE_FORCE_OAUTH` is set by the wrapper, so a host
   `ANTHROPIC_API_KEY` can never silently take over and bill the API.
 - `--ak subprocess_env_scrub=true` enables the CLI's subprocess-env scrub
@@ -158,9 +178,19 @@ Notes:
 - Subscription auth is env-only (`CLAUDE_CODE_OAUTH_TOKEN` access token,
   no refresh token, no credential file in the container); the same
   30-min-runway host refresh as the clawcodex adapter applies.
-- This openclaude snapshot's effort ladder is low|medium|high|max (no
-  xhigh), and its model metadata predates claude-opus-4-8, so it assumes
-  a conservative 128k context for compaction purposes.
+- This openclaude snapshot (0.24.0) knows claude-opus-4-8 but **not
+  claude-opus-5**, and on opus-5 it does not merely degrade — it breaks.
+  `modelSupportsAdaptiveThinking` (`typescript/src/utils/thinking.ts:159`)
+  allowlists only opus-4-8/4-7/4-6 and sonnet-4-6, then excludes anything
+  else matching `opus`, so opus-5 falls to the `budget_tokens` branch at
+  `typescript/src/services/api/claude.ts:1731` — and `budget_tokens` is
+  removed on Opus 5, i.e. **HTTP 400 on every request** (the same failure
+  the clawcodex adapter's `source=` note describes). Model metadata is
+  missing too, so the context window falls back to the 200K default
+  (`typescript/src/utils/context.ts:17`). Keep this arm on
+  `anthropic/claude-opus-4-8` until `typescript/` is refreshed to a
+  snapshot carrying opus-5.
+- Its effort ladder is low|medium|high|xhigh|max as of 0.24.0.
 
 ## Evaluate ALL terminal-bench 2.0 tasks
 
@@ -193,7 +223,7 @@ aggregate accuracy; each trial dir has the agent's stream-json log under
 # Agent kwargs
   --ak max_turns=100        # clawcodex --max-turns (default 300)
   --ak effort=high          # clawcodex --effort (low|medium|high|xhigh|max)
-                            # xhigh is model-dependent (opus-4-8 yes,
+                            # xhigh is model-dependent (opus-5/opus-4-8 yes,
                             # sonnet-4-6/opus-4-6 no → degraded to high)
   --ak version=1.2.1        # pin the clawcodex-cli PyPI version
   --ak source=git+https://github.com/agentforce314/clawcodex@main
@@ -205,6 +235,7 @@ aggregate accuracy; each trial dir has the agent's stream-json log under
 
 # Other models/providers (Harbor convention: provider/model)
   --model deepseek/deepseek-v4-pro
+  --model anthropic/claude-opus-4-8   # the previous tb2.1 baseline
   --model anthropic/claude-opus-4-5   # needs ANTHROPIC_API_KEY
 ```
 

eval/harbor/clawcodex_agent.py

@@ -39,9 +39,12 @@
   own default of 50 is too low for terminal-bench tasks). The
   ``CLAWCODEX_MAX_TURNS`` host env var works as a fallback.
 * ``effort`` — clawcodex ``--effort`` (low|medium|high|xhigh|max) for
-  models that support ``output_config.effort`` (Opus 4.6/4.8, Sonnet 4.6,
-  Fable 5). ``xhigh`` is model-dependent (opus-4-8 yes; sonnet-4-6/
-  opus-4-6 no) — clawcodex degrades it to ``high`` where rejected.
+  models that support ``output_config.effort`` (Opus 5, Opus 4.6/4.8,
+  Sonnet 4.6, Fable 5). ``xhigh`` is model-dependent (opus-5/opus-4-8 yes;
+  sonnet-4-6/opus-4-6 no) — clawcodex degrades it to ``high`` where
+  rejected. A model missing from clawcodex's effort allowlist drops the
+  flag SILENTLY, so a new model needs a clawcodex build that registers it
+  (see ``source`` below) before an effort number means anything.
   ``CLAWCODEX_EFFORT`` host env var works as a fallback.
 * ``version`` — pin a ``clawcodex-cli`` PyPI version (default: latest).
 * ``source`` — full pip-installable spec overriding the PyPI package, e.g.

src/cli.py

@@ -227,6 +227,9 @@ def main():
     return launch_ink_tui(
         provider=args.provider,
         model=args.model,
+        # --effort applies interactively too: forwarded to the agent-server
+        # child, which seeds the session's /effort level from it.
+        effort=args.effort,
         permission_mode=args._resolved_permission_mode,
         is_bypass_available=args._resolved_is_bypass_available,
         workspace=(worktree_session.worktree_path if worktree_session else None),
@@ -474,15 +477,18 @@ def _build_parser() -> argparse.ArgumentParser:
         default=None,
         help='Override the provider (anthropic, openai, zai, minimax, openrouter, deepseek, meta)',
     )
-    # Mirrors TS ``--effort <level>`` (main.tsx:995). TS registers it
-    # session-wide; the interactive TUI equivalent here is the persisted
-    # ``/effort`` setting, so the flag is exposed on the print path where
-    # no dialog exists. None = auto (settings.effort, else the parameter is
-    # omitted and the API applies its model default). xhigh acceptance is
-    # model-dependent — resolve_thinking_effort clamps it to high on models
-    # that reject it. KNOWN LIMIT: the flag governs the MAIN loop only;
-    # subagents (Agent tool) resolve from settings.effort — persist
-    # ``/effort`` (or seed settings) for session-wide coverage.
+    # Mirrors TS ``--effort <level>`` (main.tsx:995), and like TS it is
+    # session-wide: the print path reads it via HeadlessOptions, and the
+    # interactive path forwards it to the agent-server child, which seeds
+    # the session's ``/effort`` level from it (a later ``/effort`` in the
+    # TUI overrides; ``/effort auto`` clears back to settings.effort).
+    # Registered under the noninteractive group for backwards-compatible
+    # ``--help`` grouping only. None = auto (settings.effort, else the
+    # parameter is omitted and the API applies its model default). xhigh
+    # acceptance is model-dependent — resolve_thinking_effort clamps it to
+    # high on models that reject it. KNOWN LIMIT: the flag governs the MAIN
+    # loop only; subagents (Agent tool) resolve from settings.effort —
+    # persist ``/effort`` (or seed settings) for session-wide coverage.
     noninteractive.add_argument(
         '--effort',
         choices=('low', 'medium', 'high', 'xhigh', 'max'),

src/command_system/effort_command.py

@@ -102,7 +102,8 @@ def _invalid_msg(raw: str) -> str:
 
 # TS effort.tsx:179 help text. xhigh/max availability is model-dependent;
 # resolve_thinking_effort clamps xhigh to high on models that reject it
-# (wire-probed 2026-07-18: opus-4-8 accepts xhigh; sonnet-4-6/opus-4-6 400).
+# (wire-probed 2026-07-18: opus-4-8 accepts xhigh; sonnet-4-6/opus-4-6 400.
+# 2026-07-25: opus-5 accepts xhigh and max. fable-5 unprobed.)
 _USAGE = (
     "Usage: /effort [low|medium|high|xhigh|max|auto]\n\n"
     "Effort levels:\n"

src/entrypoints/agent_server_cli.py

@@ -130,6 +130,14 @@ def run_agent_server_subcommand(argv: list[str]) -> int:
                         help="Optional bearer token required on POST /sessions.")
     parser.add_argument("--provider", default=None, help="Provider name override.")
     parser.add_argument("--model", default=None, help="Model override.")
+    parser.add_argument(
+        "--effort", default=None,
+        choices=("low", "medium", "high", "xhigh", "max"),
+        help="Reasoning effort for the session. Seeds /effort; routed to "
+             "output_config.effort on Anthropic and reasoning_effort on "
+             "OpenAI-compatible providers. (_build_runtime re-validates, "
+             "because --stdio callers bypass this parser.)",
+    )
     parser.add_argument(
         "--fallback-model", default=None, dest="fallback_model",
         help="Model to switch to after repeated overloaded (529) errors "
@@ -233,6 +241,7 @@ def run_agent_server_subcommand(argv: list[str]) -> int:
     agent_config = AgentServerConfig(
         provider_name=args.provider,
         model=args.model,
+        effort=args.effort,
         fallback_model=args.fallback_model,
         permission_mode=args.permission_mode,
         is_bypass_available=is_bypass_available,

src/entrypoints/tui_launcher.py

@@ -51,6 +51,12 @@ def run_tui_launcher(argv: list[str]) -> int:
     )
     parser.add_argument("--provider", default=None)
     parser.add_argument("--model", default=None)
+    # choices= so a bad level is a clean argparse error rather than a value
+    # that silently resolves to settings.effort downstream (the server also
+    # validates the seed, since it can be driven without a parser).
+    parser.add_argument("--effort", default=None,
+                        choices=("low", "medium", "high", "xhigh", "max"),
+                        help="Reasoning effort — seeds the session's /effort level")
     parser.add_argument("--permission-mode", default="default", dest="permission_mode")
     parser.add_argument(
         "--dangerously-skip-permissions", action="store_true",
@@ -151,6 +157,7 @@ def run_tui_launcher(argv: list[str]) -> int:
     return launch_ink_tui(
         provider=args.provider,
         model=args.model,
+        effort=args.effort,
         permission_mode=mode,
         is_bypass_available=is_bypass_available,
         workspace=args.workspace,
@@ -163,6 +170,7 @@ def launch_ink_tui(
     *,
     provider: str | None = None,
     model: str | None = None,
+    effort: str | None = None,
     permission_mode: str = "default",
     is_bypass_available: bool = False,
     workspace: str | None = None,
@@ -186,6 +194,7 @@ def launch_ink_tui(
     args = SimpleNamespace(
         provider=provider,
         model=model,
+        effort=effort,
         permission_mode=permission_mode,
         is_bypass_available=is_bypass_available,
         workspace=workspace,
@@ -256,6 +265,8 @@ def _agent_server_cmd(args) -> list[str]:
         cmd += ["--provider", args.provider]
     if args.model:
         cmd += ["--model", args.model]
+    if getattr(args, "effort", None):
+        cmd += ["--effort", args.effort]
     return cmd
 
 
@@ -307,6 +318,7 @@ def _print_connect(args) -> int:
           if getattr(args, "is_bypass_available", False) else []),
         *(["--provider", args.provider] if args.provider else []),
         *(["--model", args.model] if args.model else []),
+        *(["--effort", args.effort] if getattr(args, "effort", None) else []),
         "--workspace", workspace,
     ])
 

src/models/configs.py

@@ -88,6 +88,40 @@ class ModelConfig:
         cost_cache_create_per_mtok=12.50,
         cost_cache_read_per_mtok=1.00,
     ),
+    # Claude Opus 5 — the current Opus tier, at Opus 4.8's price (5/25 per
+    # MTok) and shape: 1M context (default AND maximum), 128K true output
+    # cap, same 32_000 first-attempt wire ``max_tokens`` convention as the
+    # rows above. Registered LAST among the opus rows because its prefix
+    # base is the family-wide ``claude-opus`` (``key.rsplit("-", 1)[0]``).
+    # 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.
+    # 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
+    # window. Accepted because every Opus since 4.6 ships 1M, and it
+    # matches ``claude-fable-5``, whose base is likewise family-wide.
+    # (``src/services/pricing.py`` deliberately went the other way: its
+    # bare ``claude-opus-4`` prefix was dropped so unknown ids price as
+    # None rather than wrong. A wrong cost is silent; a wrong window is
+    # not — but the directions genuinely differ, so don't "harmonize"
+    # them without re-reading both rationales.)
+    "claude-opus-5": ModelConfig(
+        model_id="claude-opus-5",
+        display_name="Claude Opus 5",
+        context_window=1_000_000,
+        max_output_tokens=32_000,
+        supports_thinking=True,
+        supports_computer_use=True,
+        cost_input_per_mtok=5.0,
+        cost_output_per_mtok=25.0,
+        cost_cache_create_per_mtok=6.25,
+        cost_cache_read_per_mtok=0.50,
+    ),
 
     # Claude 3.7 series
     "claude-3-7-sonnet-20250219": ModelConfig(

src/providers/__init__.py

@@ -23,7 +23,10 @@ class ProviderInfo(TypedDict):
         "available_models": [
             # Frontier (above Opus tier)
             "claude-fable-5",
-            # Claude 4 series (latest)
+            # Claude 5 series (Opus tier; sonnet-5 not registered yet —
+            # it needs the same model-table entry opus-5 got)
+            "claude-opus-5",
+            # Claude 4 series
             "claude-sonnet-4-6",
             "claude-sonnet-4-5",
             "claude-sonnet-4-5-20250929",
@@ -294,6 +297,33 @@ def provider_env_vars(provider_name: str) -> tuple[str, ...]:
     return spec.env_vars if spec else ()
 
 
+def is_anthropic_wire(provider: Any) -> bool:
+    """Whether ``provider`` speaks the ANTHROPIC request shape.
+
+    The single source of truth for a decision several call sites have to
+    agree on, because the two families take the same concepts as different
+    parameters and sending the wrong one is a hard 400:
+
+    * reasoning effort — ``output_config.effort`` here vs a top-level
+      ``reasoning_effort`` body field on OpenAI-compatible providers,
+    * the system prompt — a ``system`` kwarg here vs a prepended
+      ``{"role": "system"}`` message there.
+
+    Minimax is included because it speaks the Anthropic shape. A custom
+    ``base_url`` / third-party gateway is a FLAG on AnthropicProvider
+    (``has_custom_endpoint``), never a subclass, so an isinstance test
+    cannot be fooled by one. Takes a provider INSTANCE, not a name — the
+    class is what carries the wire contract.
+
+    Imports are local: this module is imported at startup by the fast-path
+    CLI, and the provider modules pull in their SDKs.
+    """
+    from .anthropic_provider import AnthropicProvider
+    from .minimax_provider import MinimaxProvider
+
+    return isinstance(provider, (AnthropicProvider, MinimaxProvider))
+
+
 def provider_requires_api_key(provider_name: str) -> bool:
     """Whether a provider needs an API key.
 
@@ -377,6 +407,7 @@ def resolve_api_key(
     "ChatResponse",
     "get_provider_class",
     "get_provider_info",
+    "is_anthropic_wire",
     "canonical_provider_name",
     "provider_env_vars",
     "provider_has_credentials",

src/providers/anthropic_provider.py

@@ -921,7 +921,10 @@ def get_available_models(self) -> list[str]:
         return [
             # Frontier (above Opus tier)
             "claude-fable-5",
-            # Claude 4 series (latest)
+            # Claude 5 series (Opus tier; sonnet-5 not registered yet —
+            # it needs the same model-table entry opus-5 got)
+            "claude-opus-5",
+            # Claude 4 series
             "claude-sonnet-4-6",
             "claude-sonnet-4-5",
             "claude-sonnet-4-5-20250929",