v0.9.5

0.9.5 - 2026-08-12

• Refresh the pinned Node 22 image, WebSocket and Gemini SDKs, TypeScript runner, and CodeQL action. The production image now installs dependencies once and prunes development tools for the runtime layer. All compatibility, security, Linux, and Windows checks pass with this dependency set.

What's Changed

• chore(deps-dev): bump tsx from 4.23.7 to 4.23.11 in the npm-development-minor-and-patch group across 1 directory by @dependabot[bot] in #59
• chore(deps): bump node from f32b810 to d649c27 by @dependabot[bot] in #58
• chore(deps): bump ws from 8.21.2 to 8.21.3 by @dependabot[bot] in #60
• chore(deps): bump the actions-minor-and-patch group across 1 directory with 2 updates by @dependabot[bot] in #62
• chore(deps): bump @google/genai from 2.15.0 to 2.16.0 by @dependabot[bot] in #61
• Release v0.9.5 dependency refresh by @bielcarpi in #64

Full Changelog: v0.9.4...v0.9.5

v0.9.4

0.9.4 - 2026-08-11

• Test the bundled plugin and required API capabilities against the official Hermes Agent v0.20.0 image. Add scheduled and release compatibility gates. The doctor now checks the installed Hermes version.
• Add hermes-live upgrade to reconcile the npm package with the installed plugin, configuration, Hermes bridge, and user services.
• Add hermes-live diagnostics for a private, redacted support bundle. It excludes logs, prompts, task results, audio, and secret values. It refuses to overwrite a file.
• Clarify that task state, notifications, and retained results are durable, while in-progress computation still depends on the current Hermes Agent process.

What's Changed

• Release v0.9.4 compatibility and operations by @bielcarpi in #63

Full Changelog: v0.9.3...v0.9.4

v0.9.3

0.9.3 - 2026-08-11

• Start normally on Windows when the filesystem rejects directory fsync, while keeping file writes atomic and surfacing real I/O errors. Add a Windows Node 24 CI lane.
• Emit each Gemini audio part once. The adapter no longer repeats the SDK's derived top-level data value alongside its source parts.
• Raise the browser playback buffer from 5 seconds to 2 minutes. Keep a hard memory bound and show one detailed warning if a response exceeds it.
• Honor explicit OpenAI audio.end messages in push-to-talk and VAD modes. Clients that stop sending packets at turn end no longer need to synthesize silence.
• Enable OpenAI input transcription by default and forward completed user transcripts without repeating the assistant transcript. The model and optional language hint are configurable.
• Resolve the release audit advisory in the development dependency tree.

What's Changed

• Fix realtime audio reliability on Windows, Gemini, and OpenAI by @bielcarpi in #57

Full Changelog: v0.9.2...v0.9.3

v0.9.2

0.9.2 - 2026-08-05

• Make the product promise consistent across npm, the README, CLI, Dashboard, and plugin metadata: continuous voice for saved Hermes chats while durable work runs and reports back.
• Add a branded social preview and focused discovery keywords for Hermes, local speech-to-speech, and full-duplex voice directories.

What's Changed

• Clarify continuous voice positioning and release v0.9.2 by @bielcarpi in #51

Full Changelog: v0.9.1...v0.9.2

v0.9.1

0.9.1 - 2026-08-05

• Make hermes-live setup the complete install path. It now configures Hermes' private API bridge, installs the Dashboard plugin, starts both services, waits for them to become ready, and warms a real task call. Custom and remote Hermes deployments remain operator-managed.
• Run the tested Hugging Face speech-to-speech stack as a private Apple Silicon service. Setup owns its install, startup, endpoint selection, health checks, log limits, retries, and removal when switching providers; a second terminal is no longer part of the normal workflow.
• Let Hermes choose its configured model instead of sending the old hermes-agent alias as a literal model override. New and resumed chats resolve the real selection, while HERMES_MODEL remains available for deliberate overrides.
• Resolve config, plugin state, and task state from HERMES_HOME, including named Hermes profiles.
• Add protocol v6 microphone pause by voice or UI. Pausing keeps playback, the conversation, and background work alive; resuming always requires an explicit user action.
• Make Dashboard Connect wait for the gateway, Hermes, task store, and provider to be ready, then start the microphone automatically. Disconnect remains immediate even while browser permission is pending, and late microphone grants are discarded.
• Keep saved-chat conversation responsive while durable work runs. The voice can inspect bounded live activity, target tasks by a short spoken reference, start follow-ups, and report completion without exposing internal reasoning or opaque IDs in normal speech.
• Harden local provider compatibility around exact speech turns, tool-call history, fragmented transcripts, empty responses, response deadlines, disconnects, provider crashes, and the upstream single-pipeline limit.
• Preserve accepted work when the voice provider dies or disconnects. Public task state now distinguishes confirmed terminal outcomes from unknown ones instead of guessing or repeating a possibly accepted mutation.
• Detect occupied local ports and unrelated processes before setup claims them. Managed endpoints move safely when possible, every readiness probe verifies Hermes Live service identity, and normal setup leaves the standalone browser demo disabled.
• Keep local speech and assistant text out of managed runtime logs while retaining operational warnings. The package also excludes Python cache artifacts and provides configuration-free CLI help.

What's Changed

• test: make release hydration fixture deterministic by @bielcarpi in #49
• chore: release v0.9.1 by @bielcarpi in #50

Full Changelog: v0.9.0...v0.9.1

v0.8.0

0.8.0 - 2026-08-01

• Add fully local voice on Apple Silicon through the official Hugging Face speech-to-speech stack, with a pinned one-command launcher and no provider API key.
• Add a dedicated local realtime adapter and protocol v5 support for 24 kHz PCM audio, provider VAD, interruption, tool calls, and background-task updates.
• Make local voice a first-class option in setup, doctor, readiness, the Dashboard, browser client, and Docker while keeping Gemini and OpenAI support.
• Simplify the environment template, README, and supporting docs, and update the dependency and security baseline.

What's Changed

• Hermes Live Voice v0.8.0 by @bielcarpi in #45

Full Changelog: v0.7.1...v0.8.0

v0.7.1

0.7.1 - 2026-07-20

• Stop Dashboard and terminal clients from reusing fixed titles for every new Hermes chat. Fresh voice sessions can now be created repeatedly without Hermes rejecting a duplicate title.

What's Changed

• fix: allow repeated new voice chats by @bielcarpi in #42

Full Changelog: v0.7.0...v0.7.1

v0.7.0

0.7.0 - 2026-07-19

• Bind each protocol v4 voice session to a real persisted Hermes conversation. Dashboard and terminal users can start a new chat or resume an existing session, including its current writable compression tip, while protocol v3 remains available as an unbound compatibility mode.
• Keep canonical conversation turns in Hermes Sessions Chat so Hermes continues to own history, memory, tools, and skills. Long-running work stays on a separate durable Runs plane, allowing voice to remain responsive and tasks to survive client or provider disconnects.
• Add sanitized live task activity from Hermes tool lifecycle events. Authenticated owner clients can ask what a task is doing without exposing raw reasoning, full tool arguments/output, upstream run ids, or provider envelopes.
• Add durable follow-ups through the realtime provider, browser SDK, wire protocol, Dashboard, and terminal. Each follow-up is a new exclusive Hermes worker seeded with the selected terminal task's bounded retained result and linked by explicit parent/root task ids.
• Add authenticated conversation list/create endpoints and a same-origin Dashboard proxy so saved chats are usable without exposing Hermes or gateway credentials in browser code.
• Update the terminal with --resume, --unbound, and /followup, keep its default as a new persisted chat, and report the selected Hermes session after connection.
• Let restart reconciliation fall back quietly to run-status polling when Hermes has already removed a terminal run's SSE stream; the recovered result or confirmed-missing state remains authoritative.
• Tighten the public docs around the two-plane architecture and remove the speculative roadmap page. Conversation continuity is canonical; task follow-ups preserve durable work-product lineage rather than claiming hidden process resumption or automatic multi-agent orchestration.

What's Changed

• feat: continue Hermes conversations by voice in v0.7 by @bielcarpi in #41

Full Changelog: v0.6.0...v0.7.0

v0.6.0

0.6.0 - 2026-07-18

• Add hermes-live setup as the normal installation path. It securely reuses Hermes's existing API Server key when available, prompts without echoing missing provider credentials, writes an allowlisted private config, installs and enables the matching plugin, verifies Hermes capabilities and a real Gemini/OpenAI session, and waits for the gateway to become ready.
• Add a native user-service lifecycle through launchd on macOS and systemd on Linux. hermes-live service install|uninstall|start|stop|restart|status|logs keeps the gateway available without a dedicated terminal, while service definitions contain only the managed-config path and never API keys.
• Add hermes-live doctor with human and JSON output for Node compatibility, config integrity, plugin/runtime version parity, Hermes CLI and durable-run capabilities, provider configuration and optional live session, service state, and gateway readiness. Every failed check includes an exact remediation.
• Store managed settings under ~/.hermes/hermes-live/config.env with private permissions, atomic writes, process-environment precedence, an explicit key allowlist, and strict rejection of symlinks, unsafe permissions, duplicate/unexpected settings, oversized input, and non-string values. Project .env files are never loaded or executed.
• Make setup the primary README, plugin, Dashboard, and support path. The Dashboard now turns missing configuration, offline services, and failed readiness into the corresponding setup, service, and doctor commands.
• Verify the complete activation path from a clean packed npm install: Hermes key import, managed config permissions, exact plugin version, gateway startup, real readiness response, and redacted doctor output. Protocol v3 and the durable background-task contract are unchanged.

What's Changed

• feat: one-command activation for v0.6.0 by @bielcarpi in #40

Full Changelog: v0.5.0...v0.6.0

v0.5.0

0.5.0 - 2026-07-16

• Keep the realtime conversation open while Hermes works. Delegated work is now owned by a durable server-side supervisor, survives client detachment, and reports retained results through an unread completion inbox after reconnect.
• Persist a task before accepting it, reconcile known Hermes runs after a gateway restart, and fence ambiguous dispatches as dispatch_unknown instead of retrying a mutation that Hermes may already have accepted. In-progress work still cannot survive a Hermes Agent restart.
• Run work exclusively by default. Operators can opt into parallel read-only tasks with disjoint resource keys only when they explicitly trust model-declared scopes through HERMES_LIVE_TRUST_DECLARED_READ_ONLY=true.
• Add exact background-task controls to the realtime tools, terminal, Dashboard, browser client, and wire protocol v3. Speech interruption and task cancellation remain separate operations.
• Harden the local task store with a private atomic file, strict single-writer lock, bounded retention, startup reconciliation, task-store readiness checks, offline containment and abandoned-lock recovery commands, and bounded shutdown. The stable state format reads beta state but becomes forward-only after writing new recovery fields.
• Deliver completion speech through bounded, retryable session claims and mark notifications spoken only after provider handoff. OpenAI isolates each task notice from conversation history, keeps a newly finished notice behind the user's current VAD turn, and targets interruption to the exact notice response; Gemini notices remain best effort.
• Remove the provisional queuePosition hint from protocol v3 because the bounded scheduler is not a strict FIFO queue. The field was optional in beta clients, but beta gateways and clients should still be upgraded together before reconnecting to stable v0.5.
• Keep approvals fail-closed until Hermes exposes identity that can target one exact request safely. Unknown or confirmed-missing runs can be contained without fabricating a result.
• Add a real built-image Docker smoke, pin the Hermes v0.18.2 fixture to its published image digest and revision, update @google/genai to 2.12.0, and verify the packed CLI, plugin, Dashboard, browser client, task recovery, and provider adapters.
• Export the new tasks.trustDeclaredReadOnly, task-store health(), readiness tasks, and input_speech_stopped contracts. Simplify the README, Dashboard, demo, and release docs around the project's actual role: real-time voice for Hermes Agent.

What's Changed

• feat: keep voice live with durable background tasks by @bielcarpi in #38
• feat: durable real-time voice for Hermes v0.5.0 by @bielcarpi in #39

Full Changelog: v0.4.0...v0.5.0