Critical Alerts: May 2026 Bug Tracker
🔴 [P1] ChatGPT/Codex Integration Broken (#32892)
Issue: openai-codex provider crashes on every message with TypeError: 'NoneType' object is not iterable. Non-retryable — fallback chain doesn't help.
Impact: All Hermes agents configured with ChatGPT models (gpt-5.5 via codex endpoint) are completely unusable. Affects all messaging platforms.
Source: Issue #32892
🔴 [P1 Critical] Cron Jobs Completely Broken (#20586)
Issue: All cron jobs fail with provider.sort: Invalid input error in OpenRouter provider selection logic.
Impact: Major feature regression; automated scheduled tasks are non-functional on macOS and other platforms.
Source: Issue #20586
🔴 [P2] Docker Sandbox Profile Divergence (#32049)
Issue: Docker sandbox mirrors host profile state, creating silent divergent SOUL.md files. Tools report success but host state remains unchanged.
Impact: Silent failure of profile updates; agents operate on stale or conflicting "soul" state.
Source: Issue #32049
🔴 [P1] Docker Non-Root Regression — s6-overlay Boot Loop (#34648)
Issue: Regression between v0.14.0 and v0.15.0/v0.15.1. Non-root Docker users (user: "10000:10000") hit an infinite s6-overlay restart loop: s6-applyuidgid: fatal: unable to set supplementary group list: Operation not permitted. The new docker.sock group-management logic in stage2-hook.sh crashes during privilege handling.
Workaround: Roll back to v0.14.0. No other fix exists.
Impact: All non-root Docker deployments broken on v0.15.0+.
Source: Issue #34648
🔴 [P1] Conversation Compression Desynchronizes Session IDs — Silent Message Loss (#34089)
Issue: Conversation compression silently mutates the session ID mid-operation. Two anomalies observed: (A) compression started: session=A → compression done: session=B — state is written to a new session ID while the agent's bookkeeping references the original. (B) Cross-session compression aliasing — two sessions compress on each other's behalf. Result: agents emit confident "idle"/"no record" reports when they have unprocessed inbound traffic.
Real-world impact: 9-agent coordination cluster on Windows 11 (v0.14.0) observed agent-to-agent dot-loops requiring manual NSSM gateway service bounce. Inter-agent acknowledgment protocols break. Agents hallucinate self-improvement memory entries based on false "I have no memory of X" findings.
Source: Issue #34089
The Hermes Agent ecosystem is currently experiencing several high-severity infrastructure failures. For production users, these issues represent significant operational risks.
🔴 [P1 Critical] Skills Installation Failure (#30482)
Issue: The official skills index is broken. hermes skills install official/<skill> fails for all 81 official skills.
Impact: Users cannot extend agent capabilities via the standard CLI. Workaround requires manual GitHub path resolution, which bypasses versioning safety.
Source: Issue #30482
🟠 [P2] Memory Tool: Hard Character Cap Failure (#30267)
Issue: The memory tool fails when reaching hard character limits, and concurrent edits lead to entry collisions and data corruption.
Impact: Unreliable long-term state. Agents may silently lose data or overwrite critical context during high-concurrency tasks.
Source: Issue #30267
🟠 [P2] hermes update Reports "7 Commits Behind" After Switching to Release Tag (#38618)
Issue: hermes update tracks the main branch, not the latest release tag. After switching to v2026.5.29.2 (v0.15.2), the CLI still reports "7 commits behind" because the detached release tag is behind main. Users get conflicting version signals.
Impact: Users cannot trust version reporting. The CLI actively lies — says you're behind and tells you to update, but updating does nothing.
Source: Issue #38618 — opened Jun 4, 2026 by @tsktech
🟠 [P2] Windows hermes update uv Install Fails Even When uv Already Installed (#38617)
Issue: On Windows, hermes update unconditionally runs the PowerShell uv installer. When it fails (exit status 1), Hermes does NOT fall back to an existing uv already on PATH.
Impact: Windows users see scary "Failed to install managed uv" errors even though uv is already installed. Update partially succeeds via pip fallback, but the error erodes trust in the tooling.
Source: Issue #38617 — opened Jun 4, 2026 by @tsktech
🔴 [P1] Logic Failure: Repeated Tool Loops (#30408)
Issue: Agents are getting stuck in infinite or repeated tool execution loops, failing to progress while consuming excessive tokens.
Impact: High operational cost and task failure. Requires manual intervention to kill the process.
Source: Issue #30408
🔴 [P1] Telegram Integration: DM Communication Broken (#30411)
Issue: Direct Message communication via the Telegram gateway is non-functional in the latest release.
Impact: Breaks all automated workflows relying on Telegram as a primary interface.
Source: Issue #30411
The Gobii Advantage: Managed Stability
While Hermes users navigate these breaking infrastructure bugs, Gobii users benefit from a managed, pre-vetted ecosystem.
| Feature | Hermes Agent (Community) | Gobii Managed |
|---|---|---|
| Tool/Skill Reliability | Community-maintained (Subject to Index Failures) | Pre-vetted, Managed Tool Library |
| Logic Containment | Infinite Loops Possible | Built-in Loop Detection & Budgeting |
| Gateway Support | Fragile Integrations | Enterprise-grade Webhooks & Connectors |
| Update Safety | Bleeding Edge (Frequent Breaks) | Staged Rollouts & Verified Stability |
Workaround for #30482
If you must use Hermes, use the direct GitHub URL for installation until the index is patched:
hermes skills install https://github.com/NousResearch/hermes-skills/tree/main/official/<skill_name>
🔴 Tool Calling Instability & Hallucinations (#8993)
Status: Active | Severity: High
Since v0.8.0, users report frequent hallucinations where the agent claims to have found data via web_search or FireCrawl that doesn't exist, or returns empty responses. This breaks the core reliability of autonomous research loops.
🟠 Broken Skill Downloads (#5053)
Status: Active | Severity: Moderate
Official documentation skill download links are currently non-functional, preventing users from adding capabilities via the web UI as advertised. Impacts the "5-minute install" claim.
🔴 Silent SessionDB Data Loss (#2999)
Status: Confirmed | Severity: High
A critical bug in the SessionDB layer causes silent failures during session saves, leading to permanent data loss. Users may find their agent's long-term memory or session history wiped without warning.
🔴 v0.13.0 General Instability (#22315)
Status: Open | Severity: High
The latest May 2026 release (v0.13.0) is reported as "fairly unstable" by multiple users, with frequent crashes and unpredictable behavior in core agent loops.
🟠 Context Auto-Detection Failure (#24640)
Status: Confirmed | Severity: Medium-High
Hermes v0.13.0 fails to correctly auto-detect context windows for local providers (oMLX), defaulting to 256k and causing HTTP 400 errors when the actual limit (e.g., 120k) is hit.
🔴 [P1] Telegram Adapter FD Leak — Silent httpx Connection Exhaustion (#31599)
Issue: Telegram adapter's general-request pool silently accumulates hundreds of half-closed (CLOSED) sockets through HTTP proxy over ~2 days. The existing drain fix in #17015 only covers the polling pool. gateway_state.json reports connected the entire time, so external monitoring never sees the wedge.
Impact: OS fd count exceeds macOS launchd default maxfiles=256, causing all sends to fail with httpx.ConnectError: All connection attempts failed. Kanban dispatcher and channel-directory writes cascade-fail with [Errno 24] Too many open files and sqlite3.OperationalError: unable to open database file.
Workaround: hermes gateway restart clears leaked sockets (fd 287 → 54), but recurs within 1–2 days.
Suggested Fix: Bound general pool with httpx.Limits(max_connections=20, max_keepalive_connections=10) — a 2-line change at telegram.py:1414–1425. Reporter @JustinHuber offered to submit a PR.
Source: Issue #31599 | Opened: May 24, 2026 | Status: Open
🔴 [P1] Hindsight local_embedded Infinite Daemon Crash-Restart Loop When Running as Root (#13125)
Issue: When Hermes Agent runs as root on Linux with Hindsight configured in local_embedded mode, PostgreSQL's initdb explicitly refuses to run as root by design. The Hindsight daemon enters an infinite crash-restart loop that loads embedding models on every restart cycle (~958MB RAM, ~33% CPU) and degrades the entire Hermes process.
Impact: Silent resource bomb — the agent appears to "work" while Hindsight silently burns CPU and RAM. Zero user-visible feedback: no error surfaced in terminal, logs only show silent restart cycles. Many Docker/container deployments default to root.
Status: Open since April 20, 2026 — no fix after 6+ weeks.
Source: Issue #13125 | Reporter: @LehaoLin | Labels: P1, area/docker, comp/plugins, tool/memory, type/bug
🔴 [P1] delegate_task Crashes When Tool Message Content Is a List (#28639)
Issue: delegate_task(tasks=[...]) crashes with AttributeError: 'list' object has no attribute 'lstrip' when building the child tool_trace. Root cause: _looks_like_error_output() at delegate_tool.py:290 unconditionally calls .lstrip() on tool message content, but some tool messages return content as a list, not a string.
Impact: Multi-subagent delegation from Discord gateway sessions breaks entirely. The subagent fails during trace assembly, so the parent never receives results.
Status: Open. PR #33807 filed but not yet merged.
Source: Issue #28639 | Opened: May 19, 2026 | Reporter: @whiwoon