Emotion Chip v0.3.3

Bug fixes. No new features.

• Fix: the backend setting picks the same engine everywhere (#52). A value such as Fish used to send the
  speech provider's preparation down an "unknown backend" path (tagless text, ElevenLabs-style parameters) while Fish
  spoke. Case and spaces no longer matter; any other value falls back to fish, and /chip status names it with the
  fix.
• Fix: with another speech provider, the chip reports what that provider gets (#51). When Fish Audio,
  ElevenLabs or Hermes's default Edge voice speaks instead of the chip, chip_preview, chip_status, /chip status
  and the generated skill now show the dialect the chip actually renders for it (for Edge: plain text), and
  chip_preview returns exactly that text. They used to show the chip's own Fish dialect.
• Fix: a Markdown link inside a delivery span keeps its words. [whisper: See [docs](https://x.io) now.] lost
  "docs" (or turned it into a delivery tag); it now reads "See docs (https://x.io) now." in the span's delivery, as
  the same link already did outside a span.
• Docs: signals add their line to every failing tool result except a failed text_to_speech call.

Install or update:

hermes plugins install 100yenadmin/ai-emotion-chip/plugin --enable

Full notes: CHANGELOG.

Emotion Chip v0.3.2

Docs only. The plugin's code is unchanged from 0.3.1.

• Docs: the demo film plays on the front page. The README shows the demo video under the demo link.

Install or update:

hermes plugins install 100yenadmin/ai-emotion-chip/plugin --enable

Full notes: CHANGELOG.

Emotion Chip v0.3.1

Bug fixes. No new features.

• Fix: chip_preview shows what the voice is actually sent (#48). For the profile's own dialect the preview now
  runs the speech provider's own preparation for a call from Hermes. On the Fish backend that includes Fish's tag
  adapter, which turns Fish-native cues in parentheses such as (laughs) into bracket cues; the preview used to skip
  it. It also shows the final speed (Hermes's tts.speed, or Fish's speed, times the cue's), the provider's notes
  (the S1 tagless route, a pinned Fish model) and ElevenLabs style and speed settings. It still writes nothing. Any
  other dialect, and the offline hermes chip render, render the score as written.
• Docs: a front page you can act on. The README opens with what the plugin does, true badges, the install line
  that works on every Hermes (the catalog name fails on Hermes 0.23.0), three steps to a first result, voice setup for
  Fish Audio and ElevenLabs, where the agent's delivery is and is not heard, a privacy and cost summary, troubleshooting
  and how to report a bug. It links the demo, and plays one line both ways: plain, and with the agent's score.

Install: hermes plugins install 100yenadmin/ai-emotion-chip/plugin --enable (Hermes 0.22.2 or newer). Demo: https://x.com/ElectricSheepIO/status/2120173492553261309

Emotion Chip v0.3.0

The agent can see its own voice. Until now it wrote a score, called text_to_speech and learned nothing back.

• Agent tool chip_status (read-only): whether the voice can speak with this profile's settings, what to fix if
  not, the engine, model, dialect and voice in use, whether voice replies are covered, and the register the agent last
  wrote. The same facts as /chip status, from one shared source.
• Agent tool chip_preview (offline, no cost): what the voice engine would receive for a score, and which cues it
  would drop and why, without synthesizing. The same result as hermes chip render --json, with the profile's quiet
  switch applied.
• A render report on each text_to_speech result: with the chip as the speech provider, a successful result gains
  an emotion_chip entry (dialect, the agent's cue words, drops and notes, at most 600 characters), so the agent learns
  what the engine did with its score. It stays in the conversation and is sent nowhere; any media: in it is escaped so
  Hermes never reads it as an attachment.
• Both tools and the report appear only with guide_mode: generate; defer stays a byte-identical passthrough.

Install from the repository:

hermes plugins install 100yenadmin/ai-emotion-chip/plugin --ref 4f5ae2a267ebe6280e1b1bb69e8629900b7177b5 --enable

--ref takes the full commit SHA.

Emotion Chip v0.2.1

Fixes from an independent review of 0.2.0. No new features.

• Fix: the delivery log never stops speech (#43). The log is written before synthesis, and a failed write (a full
  disk, an unwritable plugin-data/ folder) used to fail every render. It now logs one warning and speech goes on;
  later rows are written once the folder is writable again.
• Fix: a failed voice reply is recovered (#44). Hermes skips its own voice note after any text_to_speech call in
  the turn, a failed one too. So when the agent's call failed (text over the model's limit, a bad argument), a chat with
  voice replies on got no voice note, or the reply read without cues. On those turns the line now asks the agent to fix
  the cause and try once more. The Hermes side is filed upstream (NousResearch/hermes-agent#151862).
• Fix: /chip status is honest about what it checked (#45). When it cannot read the profile's key, or the
  fish-audio plugin's own check fails, it says ready: probably (not checked from here: …) instead of ready: yes. A
  new voice_replies: value says whether voice replies are covered and, if not, why: no message seen yet, no
  voice-reply check in this Hermes, the wrong settings, or a restart needed after switching to the chip.
• Docs: install from the repository first. On a fresh Hermes 0.23.0, installing by catalog name can fail: Hermes
  falls back to the catalog copy it shipped with, which predates this plugin (NousResearch/hermes-agent#151856).

Install from the repository:

hermes plugins install 100yenadmin/ai-emotion-chip/plugin --ref a018d90449ddcf55363b57bd307ff2c971911065 --enable

--ref takes the full commit SHA.

Emotion Chip v0.2.0

Voice replies on messaging platforms, a hermes chip command and a readiness check.

• Voice replies carry the agent's delivery (#40). With voice replies on in a chat (/voice on, /voice tts or voice.auto_tts), Hermes used to read the reply aloud itself, without cues. Now, when a message comes in, the chip asks Hermes whether it will voice the reply. On that message's turn it adds one line asking the agent to say its answer with text_to_speech, score included, and Hermes then sends no plain voice note of its own. Nothing else changes, including the system prompt. Not on Discord, where Hermes plays its own reply in voice channels. Requires the chip as the speech provider and guide_mode: generate.
• hermes chip render and hermes chip guide (#39). See what the chip sends ElevenLabs v4, Fish S2 or Drama 3 for any score, or the guide the agent gets. Runs offline: no keys, no network.
• /chip status says whether the chip can speak (#41): ready: yes, or ready: no with each fix (the speech provider, the ElevenLabs key or voice, the Fish Audio plugin).

Try it:

hermes chip render --dialect fish_drama3 "[whisper] Ten thousand users! [pause] Cheer when the baby's [stress: awake]."

Install or update: hermes plugins install 100yenadmin/ai-emotion-chip/plugin --ref 5cff9ccd003e36494fcd9b51a17cecd37879dbfb --enable. Requires Hermes 0.22.2 or newer. Full notes in CHANGELOG.md.

Emotion Chip v0.1.1

Fixes from the first release's review backlog. No new features; the Fish model setting gains a real fish-default choice.

• Code inside a span stays text. [whisper: call items[index] now] is one whispered segment in every dialect, no longer split around [index] (#34, #35).
• /new moves /chip to the new conversation. /chip status, /chip why and /chip register follow the chat that ran /new (the plugin now registers on_session_reset). When several chats share the old session, each follows from its next message (#3, #37).
• The opt-in situation line is right on every call. With signals on, the line goes on the failing tool's own result and counts that failure (on Hermes's serial tool path it used to lag one call). A JSON result keeps its formatting and its own situation key, deeply nested JSON never raises, and nothing reads plugin state while signals is off (#3, #37).
• The skill follows a live model change. The emotion-chip-voice skill is rewritten when the Fish model or the backend changes, without a restart (#3, #37).
• fish-default. fish_model now has a real default choice: Fish picks the model by your account when nothing names one, and /chip, the delivery log and Desktop say so instead of claiming s2.1-pro (#3, #37).
• Stricter validate-data. It now rejects:
  • a sound id the parser cannot read as a sound, which would have been read as a cue, past the sound cap and /chip quiet;
  • a dialect notes value that is not a string (#4, #36).
• Docs.
  • /chip quiet is one switch for the whole profile and now says so.
  • The plugin README states that Hermes keeps only the first plugin's tool-result rewrite.
  • It also says what happens when Hermes compresses a conversation mid-reply.

Requires Hermes 0.22.2 or newer.

Install or update

hermes plugins install 100yenadmin/ai-emotion-chip/plugin --enable

Setup, settings and the full disclosure: plugin/README.md. Changes: CHANGELOG.md.

Emotion Chip v0.1.0

First release.

• The agent writes its own delivery. A small bracket score in the text of a text_to_speech call: sentence cues
  ([slightly frustrated], [warm, amused]), spans ([soft: …], [whisper: …], [slow: …], [stress: word],
  [smiling: …]), sounds ([sigh], [laugh], [chuckle], [gasp], [breath]) and pauses. The plugin never adds,
  upgrades or removes an emotion word the agent did not write.
• Two voice backends. ElevenLabs v4, v4 turbo and v3 over REST, with request stitching on v4; Fish Audio S2
  (s2.1-pro) and Drama 3 through the Fish Audio plugin in the same process. Each score is translated into that
  vendor's dialect at its documented ceiling, and anything a model cannot perform is dropped and recorded.
• Delivery guide. A system-prompt section and the emotion-chip-voice skill teach the score. Plain speech is
  always a valid choice.
• /chip command. status, why (the last render: the agent's cue words, drops and parameters), quiet and
  register.
• Optional situation facts (signals): dated task facts such as errors in a row and retries. Off by default.
• Delivery log. Local metadata without the spoken text; it can be turned off.
• CC0 emotion data. 171 concepts with attribution, rendering groups, sounds and dialect tables.
• Evaluation. Pre-registered bars in EVAL-PREREG.md; scope A3 run 3 passed its mechanical gate (#26). This
  release has no guardrails, so the guarded design's gates G1, G2 and G3 were not validly measured (amendment A3);
  run 1's observed numbers, including one G3 failure, are in evidence/run-1/.
• Hermes compatibility. Requires Hermes 0.22.2 or newer; tested on 0.23.0 (#29). The ElevenLabs key is read
  through the profile's secret scope (#30). The installable plugin lives in plugin/ (#31).

Install

hermes plugins install 100yenadmin/ai-emotion-chip/plugin --enable

Setup, settings and the full disclosure: plugin/README.md.