Registries for agents and model providers; openai-compatible provider

[mattteufel-e2b]

Why

The adapter should be easy to point at any model endpoint and, later, any agent upstream supports, without threading vendor names through the runtime. Before this PR, codex/openhands/fireworks/bedrock appeared as literals in RunOptions, _model_config, _summary_wire_api, _agent_model_id, _require_runnable_policy, _benchmark_result, _already_completed, argparse choices, preflight output, and the bundle builder.

Design

Two registries, each a frozen dataclass plus a dict:

• agents.AgentHarness (src/cybergym_e2b/agents.py): upstream --agent name; wire_api (responses | chat-completions); installs_tooling_at_runtime (drives the allowlist gate); bundle_scripts (files shipped into scripts/, e.g. the hardened Codex installer); model_id (OpenHands' openai/ prefix); optional turn_state (Codex trajectory reader that separates an interrupted turn from a graded failure).
• providers.ModelProvider (src/cybergym_e2b/providers.py): endpoint(EndpointRequest) -> Endpoint(host, base_url) and key_env names. Fireworks and Bedrock Mantle are fixed entries. openai-compatible takes --model-base-url and --model-key-env, so OpenAI, vLLM, OpenRouter, or an internal gateway work with no code change. The credential still never enters the sandbox; E2B's proxy injects it on the endpoint host only.

Everything downstream reads the registries: CLI --agent/--provider choices, preflight's provider_keys_present, _require_runnable_policy, bundle contents, _benchmark_result, batch reuse, and the experiment fingerprint (bundle_scripts_sha256 map replaces remote_install_codex_sha256).

Adding an agent upstream already runs is one AGENTS entry. Adding a fixed provider is one PROVIDERS entry plus its host in the two packaged policies (only needed for allowlist modes). README has a section on this.

Flag changes (0.4.0)

• --remote-install-codex PATH → --bundle-script NAME=PATH (repeatable, agent-agnostic)
• --bedrock-region → --model-region
• new: --provider openai-compatible --model-base-url URL --model-key-env NAME
• api.openai.com added to model_hosts in both packaged policies

Verification

• pytest -q: 62 passed (55 + 7 in tests/test_agents_providers.py)
• ruff check, ruff format --check, check_public_tree.py, actionlint, uv build, twine check pass
• Base template rebuilt from this branch resolves to the same recipe tag as the 2026-09-08 validation (cybergym-e2e-dind:recipe-3b3d5f9d827b70b3); templates.py is untouched
• Live smoke/run from this branch are the next step and will be reported separately

🤖 Generated with Claude Code

[mattteufel-e2b]

Live evidence from this branch (2026-09-15):

• templates build-base / build-ffmpeg: resolved to cybergym-e2e-dind:recipe-3b3d5f9d827b70b3 and cybergym-e2e-ffmpeg:recipe-e8010b6665ef25fa, the same tags as the 2026-09-08 validation.
• preflight ok for: smoke curl, run curl (bedrock), smoke ffmpeg, run curl (openai-compatible pointed at Fireworks). Tag-to-build receipts verified against E2B; HF access 302.
• smoke curl/arvo_66012: completed, exit 0, benchmark.status=passed, outcome=ground_truth_passed, 5.4 min, sandbox iqbvaannfvx44956xi9cf. 316 resource samples, no observation_errors; the post-workload path from PR Land PR #2 on main, remove dead code, rewrite README for customers #3 (monitor stop → summary → metrics → collect) ran cleanly for the first time on a live sandbox.