refactor(runtime): run the agent loop on apodex-agent-core 0.12.0

[zhanghanduo]

Summary

Replaces FrontierAgent's forked runtime engine with the published apodex-agent-core package (pinned ==0.12.0), following the ApodexHarness integration pattern. The core loop is now decoupled from workflows: frontier_agent.* keeps its import paths, so workflows/, apodex/ and benchmarks/ need no changes.

• 80 modules → sys.modules aliases of their agent_core counterparts (loop contracts, messages, tools, compaction, observers, AgentBus pieces, DAG, registries, models, providers). This also unifies class identity and contextvars with the engine.
• Product policy injected, not forked:
  • core/runtime/loop/agent_loop.py — AgentLoopHooks (sticky-session binding, execution scope, wall deadline, spill-file detection).
  • core/runtime/loop/tool_exec.py — ToolExecutionHooks (timeout floors, bash budget contextvar, usage metering, configurable result cap with spill, per-turn aggregate budget).
  • components/agent_bus/bus.py — configure_* resolvers so sub-agents get the same hooks, pause check and event sink.
  • infra/openai_client.py — session-affinity resolvers.
  • model-registry path, compaction prompt style and trigger ratio, ResourceManager role LLMs.
• Kept local (product-only): config, providers, llm_adapter, summary_llm, scheduler, process manager, console / duplicate-query observers.
• tools/check_symbols.py now follows sys.modules aliases and external star imports.

Behaviour changes (now follows AgentCore)

• Compaction pins the first user message verbatim and replaces legacy prose spill indexes; KeepLastN writes mini cards.
• Runaway recovery uses AgentCore's 3-stage ladder; landing turns block non-landing tool calls.
• Any-typed tool params generate {"type": "string"} (deliberate upstream: strict validators reject untyped nodes) — create_file now annotates its rows / data shapes explicitly.
• Tests that pinned fork internals (test_long_run_compaction, test_summary_retry_budget) now assert AgentCore semantics.

Upstream follow-ups (AgentCore)

Landed and released in 0.12.0, and the local workarounds are removed in this PR: ApodexAI/AgentCore#41 (empty api_key), #42 (TaskBoardObserver critical), #43 (public host helpers). Runaway-retry wording: ApodexAI/AgentCore#46. (Sub-agent result caps and fan-in protection turned out to be host-configurable already, via the AgentBus runtime spec's config_builder and TieredCompactor(protect_tool_names=...).)

Test plan

• pytest tests — 895 passed, 4 skipped (same as main)
• ruff, pyright (only pre-existing missing-extras errors), check_symbols.py, check_lazy_exports.py, import_smoke.py --stage 1
• Re-run on 0.12.0: full suite 895 passed; real react coding + agent_team smoke ✅
• Real smoke tests with apodex-1.1-mini, side by side with main (on 0.11.1):
  • react coding (fix + bash verify) ✅
  • react research (web_search / web_fetch / summary LLM) ✅
  • agent_team with parallel sub-agents via AgentBus ✅
• Long run that triggers compaction / runaway recovery with a real model
• Harbor benchmark adapter (harbor extra not installed locally)

🤖 Generated with Claude Code

[dq-ai-dev]

Reviewed current head b456dd7. The AgentCore aliases preserve shared type and context identity, while the FrontierAgent hooks retain the product-specific session, execution, timeout, metering, spill, and AgentBus policies. The focused runtime checks and current CI are green. No blocking findings.