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

[ericleepi314]

Makes clawcodex --model claude-opus-5 --effort xhigh work on a Claude subscription — interactively, not just headless. Three separate defects stood in the way.

1. claude-opus-5 was unregistered

_model_supports_extended_thinking already matches opus-5 by regex, so the loop fell through to the budget_tokens branch — removed on Opus 5, so every request 400'd. Quieter side effects: --effort was dropped (opus-5 was off the effort allowlist) and the context window resolved to the 200K default instead of 1M, the early-compaction handicap #730 removed for opus-4-8.

Registered in MODEL_CONFIGS (1M context, 5/25 per MTok), pricing.py, both provider model lists, and the three query.py gates.

Wire-probed 2026-07-25 over subscription OAuth: claude-opus-5 + adaptive thinking accepts effort: xhigh and max (200, end_turn). The xhigh entry is measured, not inferred; the "unprobed" comments now scope that to fable-5 only.

2. Interactive /effort was broken on Anthropic

• _do_set_effort validated against a hardcoded (minimal|low|medium|high) — so /effort xhigh and /effort max were rejected in the TUI while the same values worked via --effort and every other surface. Meanwhile minimal (a GPT-5 level) was accepted, and because it isn't in VALID_THINKING_EFFORT_LEVELS, resolve_thinking_effort treated it as "nothing requested" and substituted settings.effort — /effort minimal could emit max while the TUI echoed "minimal". The ladder is now VALID_EFFORT_VALUES.
• Every provider got extra_body={"reasoning_effort": …}, which is the OpenAI-compat shape. The Anthropic wire rejects it: 400 — reasoning_effort: Extra inputs are not permitted (probed). One /effort broke every following request in an interactive Anthropic session. New _turn_effort_routing routes Anthropic through thinking_effort → output_config.effort (per-model gating intact, so xhigh still clamps to high where unsupported) and keeps _EffortProvider for OpenAI-compat.

Not wrapping Anthropic providers also un-breaks the isinstance(provider, AnthropicProvider) checks that agent_loop_compat worked around and advisor.py did not.

3. --effort didn't apply interactively at all

Plumbed only into HeadlessOptions, so the interactive path parsed and discarded it. Now: cli/launcher → backend argv → AgentServerConfig → the session's /effort level, across all four entry paths (clawcodex, clawcodex tui, --print-connect, agent-server --stdio). Validated and lowercased at the seed as well as via argparse choices=, because --stdio callers reach the config without a parser.

Also

• is_anthropic_wire() in src/providers replaces a predicate copy-pasted at three sites that all had to agree.
• _EffortProvider.__getattr__ guards _inner: copy.copy built an uninitialized instance and recursed to RecursionError, which two live callers swallowed as except Exception — silently running subagents on the session model instead of their override.
• /effort and /thinking off each now warn that the other discards effort.
• The TUI advertised [minimal|low|medium|high|auto|ultracode]; it now advertises the real ladder, and the effort badge beside the model name has a producer for the first time.

Verification

• 8827 passed, 0 failed (1 deselected: test_connection_error_re_raises, a pre-existing hang confirmed on clean main).
• ui-tui tsc --noEmit clean; ui-tui tests at their 8-failed pre-existing baseline, no new failures.
• Live: headless --model claude-opus-5 --effort xhigh on subscription returns a real answer; the interactive control path accepts /effort xhigh and routes it as thinking_effort='xhigh' with the provider unwrapped.
• Every hop of the flag chain and both halves of the routing split are individually revert-sensitive — verified by deleting each line and confirming exactly the intended test fails.
• Critic-reviewed over three rounds.

Known follow-ups (not in scope here)

• claude-sonnet-5 has the identical budget_tokens 400 today — same defect class, unregistered.
• src/utils/advisor.py's advisor allowlist is opus-4-6/sonnet-4-6 only; the TS reference is two models ahead and carries @[MODEL LAUNCH] markers the port lacks.
• src/models/aliases.py's opus alias still resolves to Opus 4, so Agent(model="opus") subagents don't follow the session model.
• Interactive /effort doesn't persist to settings.effort, so subagents don't inherit it — now consistent with --effort on both paths, but a product decision worth making.

🤖 Generated with Claude Code