…ousResearch#17080)

hermes profile create --clone was copying the source profile's .env and config.yaml
verbatim, including exclusive platform credentials (TELEGRAM_BOT_TOKEN, WEIXIN_TOKEN,
DISCORD_BOT_TOKEN, etc.). These are one-to-one identity bindings — a token
belongs to exactly one bot instance and cannot be shared across profiles.

When two profiles hold the same token and start simultaneously, the platform
adapter fails during scoped lock acquisition, causing a gateway health check
timeout and a confusing 'profile loading error' in hermes-web-ui.

Fix: define _EXCLUSIVE_PLATFORM_KEYS and _EXCLUSIVE_PLATFORM_CONFIG_PATHS,
and apply _strip_env_exclusive_credentials() and
_strip_config_exclusive_credentials() to cloned .env and config.yaml files.
Exclusive credential keys are commented out in .env, and the corresponding
platform entries are force-disabled (enabled: false) in config.yaml.
A warning is printed so the user knows to configure them for the new profile.

…sResearch#17054)

Slack's app manifest rejects slash command names that start with an underscore.
_sanitize_slack_name() converts plugin command names (e.g. _reload_mcp) to
valid Slack names by lowercasing and stripping invalid chars, but the
underscore prefix was preserved. Commands like /_reload_mcp were included
in the manifest and caused Slack to reject the entire app with 'invalid name'.

Fix: skip any sanitized name that starts with an underscore before adding
it to the slash command list.

…h api_mode=anthropic_messages (NousResearch#17086)

When a custom auxiliary endpoint was configured with api_mode=anthropic_messages,
resolve_provider_client() unconditionally passed the base URL through
_to_openai_base_url(), which rewrites /anthropic → /v1. This caused all
auxiliary tasks (title_generation, compression, vision, web_extract, etc.)
to land on POST /v1/chat/completions — a 404 on Anthropic-native proxies.

Fix: add the same api_mode==anthropic_messages guard that already exists in
_try_custom_endpoint() and named-provider branches. When api_mode is
anthropic_messages, pass the URL unchanged to build_anthropic_client() and
route through AnthropicAuxiliaryClient directly.

…sResearch#17049)

Fix two locations where wmic subprocess output was decoded with
text=True (UTF-8) on Windows. Non-UTF-8 characters in process
paths cause UnicodeDecodeError, and Python ≤3.11 can return
None for stdout after decode failures causing AttributeError.
Fix: capture raw bytes, decode with utf-8 then fall back to
latin-1, and guard against None stdout.

…asoning_content (NousResearch#17052)

Changed last_reasoning extraction to only use the most recent
assistant message (current turn) rather than walking the full
message history. Previously, when the model returned null reasoning
for a simple prompt, the loop would skip the current message and
match an older assistant message's reasoning, displaying it as if
it belonged to the current response.

…tforms after 20 attempts (NousResearch#17063)

Instead of removing a platform from _failed_platforms after 20 failed
reconnect attempts, keep it in the retry queue with a fixed 10-minute
backoff. This prevents transient network/proxy outages that last hours
from permanently disabling a platform (e.g. Telegram) until the gateway
is manually restarted.

…ands (NousResearch#16843)

Make redact_sensitive_text JSON field pattern use exact key match
instead of substring match to avoid false positives like
'htpasswd_password' being treated as a credential field. Also
restrict terminal_tool redaction to structured output (JSON, logs)
only — arbitrary command output passes through unchanged so that
functional commands like htpasswd that output passwords as results
continue to work correctly.