fix(llmcore): use neutral default identity for native Claude sessions (closes #712)

[Kailigithub]

Summary

The NativeClaudeSession.raw_ask default identity prompt was hard-coded to
You are Claude Code, Anthropic's official CLI for Claude.. On machines
where self.system is empty (e.g. a freshly-copied portable directory that
fails to load mykey.py), the model retained the Claude Code identity and
introduced itself as Claude Code.

This PR swaps the unconditional default for a neutral identity and only
emits the Claude Code framing when the user has explicitly opted into the
CC-protocol relay via fake_cc_system_prompt=True. Existing
fake_cc_system_prompt=True setups are preserved bit-for-bit.

Fix

llmcore.py line 778 — the literal string is replaced with a conditional
expression:

-        payload['system'] = [{"type": "text", "text": "You are Claude Code, Anthropic's official CLI for Claude.", "cache_control": {"type": "ephemeral"}}]
+        payload['system'] = [{"type": "text", "text": "You are Claude Code, Anthropic's official CLI for Claude." if self.fake_cc_system_prompt else "You are a capable AI assistant running in GenericAgent framework.", "cache_control": {"type": "ephemeral"}}]

The downstream branches (if self.fake_cc_system_prompt: payload["system"].append(...) vs
else: payload["system"] = [...]) already condition on fake_cc_system_prompt,
so only the unconditional seed prompt needed gating.

Scope

• 1 file changed: llmcore.py
• 1 insertion, 1 deletion
• No new tests added to tests/ (the repo has no pytest suite); the
  regression harness lives at /tmp/test_issue_712.py and is documented below.

Verification

/tmp/test_issue_712.py exercises NativeClaudeSession.raw_ask with a
stubbed requests.post and asserts the outgoing system payload:

#	Scenario	system[0].text	system[1].text (if self.system)
1	fake_cc_system_prompt=False, self.system=""	neutral identity	(absent)
2	fake_cc_system_prompt=True, self.system="Custom system prompt"	Claude Code identity	Custom system prompt

Three-step dance (stash → test fails on main because the literal still
contains You are Claude Code → pop → test passes with the fix) confirms
the harness catches the regression.

$ python3 -m py_compile llmcore.py
✓ py_compile OK

$ python3 /tmp/test_issue_712.py
all identity prompt checks passed

Backward compatibility

• fake_cc_system_prompt=True (the documented path for CC relay channels
  per mykey_template.py) keeps emitting the original Claude Code
  identity text byte-for-byte.
• self.system is still prepended as before for CC relay channels, and
  used to override the default for the standard Anthropic path.
• The new default is short and avoids leaking upstream branding to portable
  users who have not configured their own system prompt.

Refs

Closes #712