Hermes Agent VS Code Extension Changelog and Release Notes

​

v0.3.2 — 2026-06-22

​

Added

• Hermes initialization hint. If a prompt produces no output for 25 seconds, the status bar displays a message confirming that Hermes is still initializing. First replies can take 1–3 minutes while the plugin stack loads — the hint keeps you from assuming the connection is broken.

​

Fixed

• Connection stuck on Connecting… / Connection error. Stream output is now accepted as soon as prompting starts. Superseded or cancelled prompts no longer leave the client in a blocked state, eliminating the most common cause of the “stuck Connecting” symptom.
• Cross-session stream bleed. Session updates whose sessionId does not match the active session are silently ignored. Messages from a background session can no longer appear in the foreground chat.
• More robust content parsing. Text extraction from ACP content blocks is now more reliable, fixing cases where partial or structured content blocks produced garbled output.

​

v0.3.1 — 2026-06-22

​

Added

• Environment detection (L0–L5). The wrench menu exposes a layered Hermes discovery scan that checks hermes --version, verifies ACP readiness, and shows compact percentage progress in the chat toolbar with expandable step details.
• ACP readiness check. After Hermes is located, the extension runs hermes acp --check. If the check fails, it automatically attempts pip install agent-client-protocol==0.9.0 and re-checks before reporting an error.
• Environment configuration. A new wrench-menu entry lets you apply the detected Hermes installation to extension settings or the system PATH. The option is always available, even when Hermes is already configured.
• Smart detection trigger. On connect, if Hermes is already configured in the extension or on the system PATH and passes a quick verification (including ACP), the full layered scan is skipped — keeping startup fast.
• Cancel detection. Closing the progress bar during an in-progress environment scan now cleanly aborts the operation.

​

Fixed

• Cursor — Settings. The More → Settings menu item now opens the VS Code Settings UI filtered to this extension. Previously it sometimes opened settings.json directly, which had no visible effect in Cursor.
• Windows PATH corruption. Writing to the user PATH on Windows no longer corrupts existing entries due to JSON/PowerShell escaping bugs.
• Detection summary and step titles now reflect the final completion state; completed steps show success icons instead of spinning indicators.

​

v0.3.0 — 2026-06-21

​

Added

• Session memory attachment. After an agent reset caused by a model or session switch, you can optionally attach prior messages as reference text to the next send. Choose the last 2 messages, last 10, all, or a custom selection — a preview and token estimate are shown before you commit.
• hermes.contextAttachVisibility setting. Controls when the memory picker appears: onNewSession (default — shown after a reset until the first successful reply), always, or never.
• FAQ modal. A new More → FAQ menu item opens a modal covering session reset behaviour, model and session switching, profile changes, model list completeness, updates, and bug reporting.
• Session reset divider. A visual divider in the chat UI marks the point where messages were restored from local history. Restored messages are read-only from the agent’s perspective — the agent does not retain that context unless Hermes itself supports session restore.
• Session switch confirmation. If you switch to a different session tab while a reply is actively streaming, the extension now asks you to confirm before interrupting the current response.

​

Fixed

• The model list now falls back to cached ACP options when a fresh fetch returns an empty list, preventing the dropdown from going blank after a reconnect.

​

v0.2.6 — 2026-06-21

​

Added

• Profile auto-discovery. When hermes.agents is not configured, the extension runs hermes profile list to discover your Hermes profiles and surfaces them as a quick-switch selector in the toolbar — no manual ID juggling required.
• Profile-isolated storage. Session history and model preferences are now stored per profile, so switching profiles gives you a clean slate for both chat history and model selection.
• Hermes profile CLI arguments, ACP model catalog parsing, and grouped model display.

​

Fixed

• Selecting the Default profile now explicitly passes --profile default to the Hermes subprocess. Previously, launching without a profile argument followed Hermes’s globally active profile, which caused the wrong default model to be shown.
• The model list is now fetched only via ACP: the extension prefers model.options from Hermes and falls back to models.availableModels from the session. Reading config.yaml directly has been removed.
• The model dropdown displays models grouped by provider, and the selected model correctly aligns with the profile’s default as returned by ACP.
• On reconnect and retry, the current profile shown in the chat UI is preserved instead of being reset from workspace settings.

​

v0.2.5 — 2026-06-21

​

Added

• In-chat permission approval cards. Permission requests from Hermes are now rendered as cards directly inside the WebView instead of appearing as VS Code warning message pop-ups. Each card offers approve/deny choices with session-level or permanent scope. Details are collapsed by default and expand when the content exceeds three lines.
• Approval history persistence. Permission cards are written to the session message history, so they survive a WebView refresh or session switch as read-only records.
• MCP config forwarding. The extension reads MCP server configurations from ~/.cursor/mcp.json and from workspace-level .cursor/mcp.json and .vscode/mcp.json, then forwards them to Hermes on session/new.
• Smart scroll during streaming. The chat view follows streaming output to the bottom by default. Scrolling manually pauses auto-scroll; it resumes automatically after 5 seconds of inactivity while a reply is still streaming.
• TOKEN ring percentage. The token usage ring in the toolbar now displays your current usage as a percentage in its centre.

​

Fixed

• After an approval card interaction or a tool call, subsequent assistant replies correctly start a new message segment instead of incorrectly appending to an existing bubble.
• The allow_session permission option no longer incorrectly displays as “Always allow” — optionId now takes priority over the kind mapping when determining label text.