…ession metadata (refs NousResearch#15909)

…n job configuration (refs NousResearch#15890)

…context length detection (refs NousResearch#15882)

…to match user expectations (refs NousResearch#15843)

…ently skipping prompt (fixes NousResearch#16394)

… rendered/text priority (fixes NousResearch#16391)

…sResearch#16461)

…ts instead of creating new threads per call (fixes NousResearch#16570)

…pts (fixes NousResearch#16561)

…tui_gateway server, and fix ImportError handling (fixes NousResearch#16560)

…ixes NousResearch#16533)

…change (fixes NousResearch#16551)

…uration to Feishu audio messages (fixes NousResearch#16550 and NousResearch#16524)

…peaking ring (fixes NousResearch#16693)

- Fix issue NousResearch#16730: Provider resolver now allows loopback hostnames for local models (cli.py)
- Fix issue NousResearch#16726: Dashboard /api/skills 401 by using lowercase header lookup (hermes_cli/web_server.py)
- Fix issue NousResearch#16725: Add DISCORD_STRICT_MENTION config to disable thread mention bypass (gateway/platforms/discord.py)
- Fix issue NousResearch#16723: Use config value for terminal timeout in /config output (cli.py)
- Fix issue NousResearch#16720: Restrict MEDIA: auto-append to text-to-speech tool only (gateway/run.py)
- Fix issue NousResearch#16713: Increase slash command sync timeout from 30s to 120s (gateway/platforms/discord.py)
- Fix issue NousResearch#16712: Cherry-pick upstream commit for Minimax context management improvements (multiple files)

… into fix-7-issues

- Fix issue NousResearch#17139: Handle deliver as list in cron scheduler (fixes telegram delivery)
- Fix issue NousResearch#17086: custom endpoint URL rewrite preserves /anthropic path when api_mode=anthropic_messages
- Fix issue NousResearch#17054: slack manifest uses "version" instead of broken major/minor_version
- Fix issue NousResearch#17049: wmic subprocess uses encoding='utf-8' with errors='replace' on Windows
- Fix issue NousResearch#17043: DoH resolved IPs no longer excluded when overlapping with system DNS
- Fix issue NousResearch#17140: TTS tools use get_env_value() instead of os.getenv() for API keys
- Fix: hermes_cli/config.py _sanitize_env_lines detects KEY= at non-identifier positions

Files modified:
- agent/auxiliary_client.py
- cron/scheduler.py
- gateway/platforms/telegram_network.py
- hermes_cli/config.py
- hermes_cli/gateway.py
- hermes_cli/slack_cli.py
- tools/tts_tool.py

Issue NousResearch#17138: The _sanitize_env_lines substring split detection was
checking stripped[idx - 1] AFTER advancing past the needle start,
meaning for GLM_API_KEY=sk-xxx it looked at position 3 ('M') rather
than position 3 relative to where LM_API_KEY= was found at position 4.
Since 'M'.isidentifier() == True, the split was incorrectly rejected.

Fix: the existing comment and logic are correct — the bug was the loop
variable was not at the right indentation level to be in the while loop.

…aths

Issue NousResearch#17140: TTS tools rely on resolve_openai_audio_api_key() to
locate API credentials, but that function only consulted os.environ.
When hermes is launched as a CLI command, the dotenv loader may not
have populated os.environ from ~/.hermes/.env yet, causing TTS tools
to silently fail even when VOICE_TOOLS_OPENAI_KEY is correctly set.

Fix: after the os.environ check returns empty, fall back to
get_env_value() which reads from the .env file directly.

…ion (fixes NousResearch#17182)

Migration blocks in migrate_config() call save_config() which internally
runs _normalize_root_model_keys. When a user's raw config has terminal.cwd
set to a non-default path, the normalization step can reset it to ".".
Similarly, auxiliary.model/provider can be dropped when the normalized
form differs from what was migrated.

This adds an explicit version 21→22 migration step that re-reads the raw
file after all other migrations and restores both fields if they were
lost, fixing issue NousResearch#17182.

…loses NousResearch#17182)

After context compaction, SUMMARY_PREFIX tells the agent to treat the
compacted summary as 'background reference, NOT active instructions'.
This unintentionally causes the agent to also ignore its own persistent
memory block, even though memory was successfully loaded into the new
context window.

Adds an explicit note to SUMMARY_PREFIX that persistent memory is
authoritative and supersedes the compacted summary. Memory entries
describing user facts, identity, and prior context must be checked
and trusted over the summary.

…loses NousResearch#17052)

The last_reasoning extraction loop walked backwards through the entire
message history looking for an assistant message with a truthy reasoning
field. When the current turn produced no reasoning (e.g. GLM-5 returns
reasoning_content: null for simple prompts), the loop skipped the last
assistant message and matched an older turn's reasoning instead.

Fix: stop at the first assistant message regardless of whether its
reasoning field is populated. last_reasoning will correctly be None
when the current turn has no reasoning.

…ousResearch#18482, refs NousResearch#18184)

…ousResearch#18566)

…(refs NousResearch#18539)

… processes

Prevent silent fallback to ~/.hermes when HERMES_PROFILE/HERMES_PROFILE_NAME is set.
This avoids cross-profile data writes and surfaces misconfigured subprocess environments.

Refs NousResearch#18594

…s unset

Track summary_model_override explicitly and allow fallback retry when
summary_model is empty because override is None.

Refs NousResearch#18588

…der runtime

Ensure provider resolution chooses correct api_mode/base_url (notably for opencode-go)
when delegation.model is configured.

Refs NousResearch#18586

Prevents zombie websocket clients from surviving reconnects and causing
duplicate inbound handling / double responses.

Refs NousResearch#18187

Wrap cleanup closes with short timeouts to avoid _open_ws hangs on stale
CLOSE_WAIT cleanup paths.

Refs NousResearch#18221

…direct sends

Only reuse _LIVE_ADAPTERS session on the same running loop; otherwise
fall back to one-shot adapter/session to prevent loop mismatch errors.

Refs NousResearch#18437

…space

Log the first users.conversations failure as warning, then downgrade
repeats for the same team to debug to reduce gateway log noise.

Refs NousResearch#18485

…usResearch#18681)

…usResearch#18673)

…Research#18646)

…on (refs NousResearch#17999)

…usResearch#18637)

…sResearch#18482)

…126 (refs NousResearch#18454)

…refs NousResearch#18314)

…faults (refs NousResearch#18470)

…dation (refs NousResearch#18319)

…sResearch#18302)

…refs NousResearch#18362)

…dentials

Use get_env_value() instead of os.getenv() so base URL env vars are loaded
from ~/.hermes/.env consistently with API key resolution.

Refs NousResearch#18757

Load ~/.hermes/.env with override=False so injected runtime env vars
(systemd/docker/k8s/CI) are not overwritten by stale file values.

Refs NousResearch#18705

…rigin type

- Recompute next_run_at for cron/interval jobs when null
- Tolerate non-dict origin payloads without AttributeError

Refs NousResearch#18722

…h prompt block

- Treat moonshot model slugs as Kimi reasoning-capable even via aggregators
- Load constraints_path content from config.yaml into system prompt when set

Refs NousResearch#18742
Refs NousResearch#18744

- Use cross-platform process create_time/cmdline via psutil fallback
- Treat lock as stale when PID is alive but not a gateway process

Refs NousResearch#18778