Switch from Bedrock to Anthropic endpoint as default. Include support for gpt-5.5

[lewtun]

Summary

When running ml-intern with the default config, the CLI was using bedrock/us.anthropic.claude-opus-4-6-v1. On machines without permission to invoke that Bedrock inference profile, even a trivial prompt failed with a litellm.APIConnectionError wrapping a Bedrock authorization error for bedrock:InvokeModelWithResponseStream.

This PR switches the default path away from Bedrock, adds direct OpenAI GPT-5 model support to the suggested model list and effort validation, and fixes a probe edge case that showed up while testing GPT-5.5.

What changed

• switch the default model from Bedrock to direct Anthropic (anthropic/claude-opus-4-6)
• update /model suggestions to surface direct Anthropic and OpenAI options
• add direct OpenAI suggestions for openai/gpt-5.4 and openai/gpt-5.5
• widen OpenAI effort validation so xhigh is accepted for GPT-5.4 / GPT-5.5
• raise the effort probe token budget so GPT-5 models do not fail capability detection with a false output-limit error
• document OPENAI_API_KEY usage in the README

Testing

Live smoke tests were run against:

• anthropic/claude-opus-4-6
• anthropic/claude-opus-4-7
• openai/gpt-5.4
• openai/gpt-5.5

Results:

• Opus 4.6 and 4.7 both accepted max reasoning effort and returned a successful response in streaming and non-streaming modes
• GPT-5.4 and GPT-5.5 both correctly fell back from max to xhigh and returned a successful response in streaming and non-streaming modes
• unit tests: uv run pytest tests/unit/test_llm_params.py

cc @jagwar for viz

[claude]

PR Review in Progress

• Gather context and read changed files
• Analyze diff and review code
• Post findings

View job run

[claude]

Claude finished @lewtun's task in 2m 14s —— View job

---

PR Review

No blocking issues — 1 P1
Verdict: ready to merge

What I checked

• Default model change: bedrock/us.anthropic.claude-opus-4-6-v1 → anthropic/claude-opus-4-6 in configs/main_agent_config.json
• _OPENAI_EFFORTS expansion to include xhigh and cascade correctness end-to-end
• GPT-5.5 / GPT-5.4 routing through _resolve_llm_params (the openai/ branch, llm_params.py:165-175)
• _PROBE_MAX_TOKENS bump from 16 → 64 and its effect on the probe cascade
• pyproject.toml rename (hf-agent → ml-intern) — checked for residual references in CI, Dockerfile, and Python imports; none break
• New tests in tests/unit/test_llm_params.py
• uv.lock — skipped per policy

---

P1 — Stale user-facing string and comment now misrepresent xhigh as Anthropic-only

The PR adds "xhigh" to _OPENAI_EFFORTS (llm_params.py:75), correctly allowing it to be forwarded to OpenAI models. But two places still tell users (and future readers) that xhigh is Anthropic-only:

agent/main.py:774 — rendered directly in the terminal when a user runs /effort:

'max' and 'xhigh' are Anthropic-only; the cascade falls back to whatever the model actually accepts.

agent/core/effort_probe.py:35 — code comment above _EFFORT_CASCADE:

# try. First success wins.  ``max`` / ``xhigh`` are Anthropic-only; providers
# that don't accept them raise ``UnsupportedEffortError`` synchronously ...

After this PR xhigh is also offered to openai/ models, so both statements are factually wrong. The cascade still works correctly regardless (it falls back if the API rejects the level), but a user running /effort xhigh after switching to openai/gpt-5.5 will see help text that tells them it won't work when it actually will. Fix this →

---

Notes

• The _PROBE_MAX_TOKENS 16 → 64 bump is well-motivated: frontier reasoning models can exhaust 16 tokens before completing even a trivial reply, causing a false "output limit reached" misclassification in the probe.
• Tests: test_openai_max_effort_is_still_rejected uses a bare try/except + raise AssertionError instead of pytest.raises() (P2 style nit, not flagging formally).
• Package rename (hf-agent → ml-intern): the only residual hf-agent references are in scripts/build_kpis.py ASCII-art comments (HF Hub dataset names, not Python package references) and frontend/package.json (a separate npm package). Neither is broken by this change.

[claude]

Claude finished @lewtun's task in 46s —— View job

---

Re-review (follow-up commit 623b5af)

No blocking issues
Verdict: ready to merge

P1 fix acknowledged — The stale xhigh messaging flagged in the previous review has been addressed in both locations:

• agent/main.py:774 now reads 'max' is Anthropic-only; 'xhigh' is also supported by current OpenAI GPT-5 models.
• agent/core/effort_probe.py:35 comment updated to match.

Both fixes are accurate and consistent with the _OPENAI_EFFORTS expansion in llm_params.py:75. No new P0 findings in this commit.

---