fix(agent): deepcopy agent.tools before llama.cpp grammar-retry sanitization

[AhmetArif0]

Root Cause

strip_pattern_and_format() is documented to mutate its input in place:

Callers that need to preserve the original should deep-copy first.
— tools/schema_sanitizer.py

The llama.cpp grammar-retry path in conversation_loop.py called it directly
on agent.tools without a deep-copy, permanently stripping pattern/format
keywords from the shared tool-registry dicts — not only for the immediate
retry, but for every subsequent API call in the same session.

Affected sessions: any that mix a llama.cpp primary model with cloud-provider
auxiliary tasks (Anthropic, OpenAI, OpenRouter, Gemini), or that switch models
mid-session after a grammar error. Those follow-on calls lose the schema hints
that cloud providers rely on for date/phone/email parameters.

Missed sibling of #34416

PR #34416 applied the identical deep-copy fix to
chat_completion_helpers.py (line 623) and auxiliary_client.py (line 708)
after identifying that the xAI sanitizer paths suffered the same permanent
mutation. This call site in conversation_loop.py was not included.

Fix

# before
_, _stripped = strip_pattern_and_format(agent.tools)

# after
_tools_copy = _copy.deepcopy(agent.tools)
_, _stripped = strip_pattern_and_format(_tools_copy)
agent.tools = _tools_copy

• Registry dicts are preserved; a future /reload-mcp or tool rebuild returns
  full schemas.
• The session-level permanent strip (so subsequent llama.cpp calls don't
  re-trigger the retry) is retained via the explicit agent.tools = _tools_copy
  assignment — behaviour is unchanged for pure llama.cpp sessions.
• The exception path is safer: a partial failure in strip_pattern_and_format
  no longer leaves agent.tools in a partially-mutated state.

Test plan

• Start a session with a llama.cpp OAI-compatible server that rejects
  pattern/format in tool schemas (HTTP 400 grammar error).
• Confirm the grammar-retry fires and the session continues.
• Switch to a cloud provider (or trigger an auxiliary task routed to one)
  in the same session — verify the tool schemas sent to the cloud provider
  still contain pattern/format keywords.
• Existing test_run_agent_codex_responses.py and schema-sanitizer unit
  tests pass unchanged.

[alt-glitch]

Missed sibling of #34416 — same deepcopy fix for strip_pattern_and_format() mutation, this time in the conversation_loop.py llama.cpp grammar-retry path (the other two call sites were fixed in #34416).

[teknium1]

Thanks for isolating the remaining in-place sanitizer call. Current main still invokes strip_pattern_and_format(agent.tools) directly at agent/conversation_loop.py:2946, while the helper documents its in-place contract at tools/schema_sanitizer.py:382-388. This is particularly consequential because the quiet-mode tool-definition cache shares schema-dict references (model_tools.py:327-335).

Problems

• The PR has no regression coverage for this recovery branch. Current tests cover classifier recognition and sanitizer behavior, but not preservation of the source schema after llama.cpp recovery.

Suggested changes

• Add a focused test that drives the grammar recovery, confirms the retry schema is stripped, and confirms the original/cached tool schema retains pattern and format.

The code change itself matches the prior deepcopy pattern in 1386a7e4789c9b886395804e8475a4252217e4ac; current surrounding code has moved to agent/conversation_loop.py, so salvage should be a small conflict-resolution cherry-pick. Automated hermes-sweeper review.

[mvdbastos]

Flagging for visibility: we're pulling #67349 into our fork, which touches this same strip_pattern_and_format(agent.tools) call site in conversation_loop.py. Worth landing this deepcopy fix alongside/before it so the two don't conflict — #67349's retry path would otherwise inherit the same shared-dict mutation bug this PR fixes. Happy to help rebase either one on top of the other if useful.