fix(agent): unwrap pre-wrapped context-engine tool schemas

[arminanton]

What does this PR do?

agent_init defensively unwraps an already-enveloped context-engine tool schema before re-wrapping it, so a misbehaving engine can't brick every turn with an HTTP 400.

A ContextEngine.get_tool_schemas() is contracted to return bare schemas ({name, description, parameters}); the host wraps each one in the OpenAI envelope ({"type": "function", "function": {...}}) before adding it to agent.tools. If an engine mistakenly pre-wraps its schemas in that envelope, wrapping again yields {"function": {"function": {...}}} whose outer function.name is empty. The provider then rejects the entire request with tools[N].function.name: empty string, so every turn fails for the whole session once that engine is active.

The fix detects the envelope and unwraps to the inner bare schema before processing, so both bare and pre-wrapped inputs register with a correct name.

Related Issue

Same error signature as #11411 (empty string name → provider 400), but a different root cause and code path: #11411 is about empty-name function_call_output on the Responses-API input side; this is the outgoing tool-schema registration in agent_init. The empty-name 400 has multiple distinct sources; this PR closes the context-engine one.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/agent_init.py — new module-level helper _normalize_context_engine_schema() that returns the inner bare schema when the input is an OpenAI envelope (type == "function" + a function dict + no top-level name), and returns bare/non-dict inputs unchanged. The context-engine tool-collection loop now calls it before wrapping, replacing the previous inline _schema.get("name", "") with a normalize-then-read.
• tests/agent/test_context_engine_tool_schema_unwrap.py — 8 new tests exercising the real helper from agent_init (not a re-implementation): bare passthrough, envelope unwrap, non-dict input, the name+type-but-not-envelope edge case, and the resulting registration shape (bare / pre-wrapped / mixed / no-empty-names-emitted).

How to Test

1. Have a context engine return a pre-wrapped schema: {"type": "function", "function": {"name": "grep", ...}}.
2. Initialize an agent with that engine active.
3. Before: the tool registers as {"function": {"function": {...}}} with an empty outer name → the next request 400s with tools[N].function.name: empty string. After: the schema is unwrapped, the tool registers with function.name == "grep", and the request is valid.

pytest tests/agent/test_context_engine_tool_schema_unwrap.py -q

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(agent):)
• I searched for existing PRs to make sure this isn't a duplicate (none touch the context-engine schema registration path; [Bug]: custom GPT-5.4 gateway sessions hit Responses API 400 "Invalid 'input[n].name': empty string" even after reset/update #11411 shares the error string but is the Responses-API input path)
• My PR contains only changes related to this fix (no unrelated commits)
• I've run pytest tests/ -q and all tests pass (8 new + the existing 29 context-engine tests in tests/agent/test_context_engine.py + test_context_engine_host_contract.py, no regression)
• I've added tests for my changes
• I've tested on my platform: Ubuntu Linux (Python 3.11)

Documentation & Housekeeping

• I've updated relevant documentation (helper docstring explaining the envelope/unwrap contract) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — N/A (no config keys)
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) — N/A (pure Python, no OS primitives; passes the Windows-footgun check)
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

[teknium1]

Thanks for the focused regression analysis.

Automated hermes-sweeper review found that current main already provides this exact protection:

• agent/memory_manager.py:49 normalizes OpenAI-enveloped schemas to their inner bare function schema and rejects nameless inputs.
• agent/agent_init.py:1939-1958 applies that normalizer to every context-engine schema before adding its OpenAI envelope.
• Commit 710cd48fb1d649925e159aa08d29993f828c40ee (fix(agent): validate context/memory tool schemas before wrapping) implemented this behavior and is contained in v2026.7.1.

Closing as implemented on main.