hermes atlas
318·repos hermes·v0.23.1 ★ star this repo

DanielLi202/hermes-tag

Post images, add notes, let the thread run, then @ it — Hermes Tag pulls the few messages that matter (originals + your notes + relevant replies), not your last line, not your whole history. Claude-Tag-style context-selection for Hermes on Feishu/Lark & Slack — @-only, no full-history RAG, per-chat memory + audit.

★ 2 langPython licenseMIT updated2026-11-03

Hermes Tag is a context-selection layer for Hermes on Feishu/Lark and Slack, with reduced capability on DingTalk. It responds when mentioned, selecting relevant earlier messages and images rather than using only the trigger message or the full transcript; it also provides per-chat memory and admin controls for retained data.

  • Selects bounded context across messages and images
  • Scopes long-term memory to each chat
  • Provides redacted audit events and chat data controls
full readme from github

Hermes Tag

Post the images, add your notes, let the thread run — then @ it. It already has everything that matters.

Use Hermes to bring Claude-Tag-style capability to your Feishu/Lark (and Slack). Drop a few images, add a note, let other people chime in — when you finally @-mention it, it pulls those images (the originals), your note, and the relevant replies — not just your last line, and never your whole history. Silent until mentioned; then it retrieves exactly the few messages that matter.

English · 中文

License: MIT Version: 0.5.0 hermes-agent: tested against v0.22.5 Python: 3.14

In a Feishu/Lark group you post three charts, add a few notes, drop an unrelated 'Lunch at 12?' line, then @ Hermes Tag last — it answers in-thread with numbered key takeaways drawn from the charts' originals, your notes, and the relevant discussion, and flags the unrelated line as not a conclusion.

Real screen recording · HD MP4

What it is

Hermes Tag is a Claude-Tag-style context-selection layer for Hermes on Feishu/Lark (and Slack). It overrides Hermes's built-in Feishu platform — it is not a new platform. Each enabled chat gets one shared agent identity, and it answers only when @-mentioned.

What it uniquely adds on top of Hermes's built-in channels:

  1. Late @, full context. You can scatter context before you ever invoke it — images, a note, other people's replies — and when you finally @-mention it, it reconstructs the right bounded evidence across time and media: the images' originals, your note, and the relevant discussion. Not just your trigger message; not the whole transcript.
  2. Per-chat memory, by design. Long-term memory is built only from @-mention interactions and stays scoped to that chat — one chat's memory never leaks into another, and never becomes your whole account history. There is no cross-channel "workspace memory"; that isolation is the privacy promise.
  3. Auditable, with redacted audit events. /tag admin audit returns redacted events (scope, time, counts — never message text); /tag admin clear|disable removes a chat's retained data. The enabled_chats allowlist is the only storage and processing boundary.

The ContextSelector chooses bounded evidence with focused_reply, thread, deictic_recent, and plain scopes instead of dumping the transcript. That means no full-history RAG and no ambient auto-answering.

Shipped now vs. roadmap. Shipped: bounded multimodal evidence, Tier-0/Tier-1 memory, per-chat isolation, admin lifecycle, and redacted audit — on Feishu/Lark and Slack. DingTalk is also supported, but with reduced capability: DingTalk bots only receive messages that @-mention them in groups (no equivalent of Feishu's im:message.group_msg), so ambient group context (Tier-0) isn't available there — see docs/dingtalk.md. Roadmap: deeper connector/source-binding parity. The Claude-Tag comparison is the goal we measure against, not a claim that every Claude-Tag feature already ships.

This repository, the Python/pip package, and the manifest name are all hermes-tag.

Security & Risk Warnings (Read Before Use)

  • The enabled_chats allowlist is the storage and processing boundary.

  • Tier-0 stores message text. Each message the bot receives in an enabled chat is kept as Tier-0 short-term context, including its text. On Feishu this happens only when granted_scopes includes im:message.group_msg; on Slack and DingTalk it always happens. Defaults: 500 rows per chat (tier0_max_count). Rows older than 24 hours (tier0_ttl_seconds) are deleted the next time a message is stored for that chat, so a quiet chat can keep them longer.

  • Tier-1 stores the question and the reply. Each @-mention turn writes a Tier-1 memory of the form question=<message text>; conclusion=<reply text>. Every turn in a DM counts as a mention. Once a chat passes tier1_max_count (100 by default), older memories are merged. Tier-1 has no time limit.

  • Both are kept in a plaintext SQLite file (declared encryption_posture: plaintext-db-on-local-disk), at <hermes home>/plugin-data/hermes-tag/<platform>-tag.sqlite3 or your db_path. Attachments on stored messages are copied into the sibling <sqlite file>.media directory. Media from quoted messages is downloaded through the bundled adapter into Hermes's media cache, one new file per download. The plugin deletes all of these files: after the turn, or on clear or eviction if they are kept as Tier-0 context.

  • On Feishu it makes these Feishu Open API calls of its own, with the bot's credentials:

    • application.v6.scope.list, once per process, to check whether the app really holds im:message.group_msg as granted_scopes claims;
    • im.v1.message.get plus message-resource downloads, for quoted messages;
    • a DM to each configured admin when that check finds a mismatch or an available upgrade.

    On Slack and DingTalk it makes no API calls of its own.

  • /tag standing confirm (admin-only) creates a Hermes cron job. The job runs on its schedule and posts to the chat until it is cancelled or paused.

  • /tag admin clear and /tag admin disable delete a chat's Tier-0, Tier-1 and media, and cancel the cron jobs behind its standing jobs. They do not reset the Hermes conversation; send /new for that.

  • Audit events (/tag admin audit) record startup, storage, admin, standing-job and lifecycle actions: ids, counts and fixed strings, never message text. They are kept after a clear. Rows written by earlier versions that held a context preview are scrubbed when the store opens.

Read SECURITY.md and docs/design/ before enabling a pilot chat.

Requirements / Compatibility

Item Requirement
hermes-agent >=0.22.5 (tested: v0.22.5)
lark-oapi 1.6.8
Python 3.14
Required env FEISHU_APP_ID + FEISHU_APP_SECRET

Hermes enforces the manifest requirement requires_hermes: ">=0.22.5".

Quickstart (<60s)

Installing with an AI agent? Point it at llms.txt and AGENTS.md — those are written for agent-driven install/config. This README is for humans.

hermes plugins install DanielLi202/hermes-tag
plugins:
  enabled:
    - hermes-tag
platforms:
  feishu:
    require_mention: false   # so unmentioned group messages reach the adapter for Tier-0 context
    extra:
      feishu_tag:
        enabled: true
        enabled_chats: [oc_xxx_pilot_chat]
        bot_open_id: ou_xxx_bot_open_id
        granted_scopes: [im:message.group_msg]
        admins: [ou_xxx_admin_open_id]
        encryption_posture: plaintext-db-on-local-disk

bot_open_id and admins must be Feishu open_id values (ou_...). This plugin does not need contact:user.employee_id:readonly; it normalizes sender identity to open_id even when that scope is granted.

On first use, Hermes Tag live-verifies granted_scopes against the Feishu app and DMs configured admins if the config claims im:message.group_msg but the app does not actually hold it. That catches the silent failure mode where non-@ group messages are never delivered, so Tier-0 group context stays @-only.

Full onboarding + live verification: see after-install.md. Slack setup: see docs/slack-setup.md. DingTalk setup + capability limits: see docs/dingtalk.md. Hitting something that looks broken? Check docs/known-limits.md — it maps common symptoms to Feishu-platform / hermes-core limits this plugin can't fix.

Usage

In groups, /tag commands require @-mention.

  • /tag status -> admin-only; reports the adapter platform, capability flags (tier0_full_ingest / l2_context / tier1_at_memory), and current per-chat metrics.
  • /tag admin count -> returns retained-state counts such as tier0=... tier1=... standing_jobs=....
  • /tag admin clear|disable|audit -> clears plugin data, disables plugin storage for the chat, or returns redacted audit events.
  • /tag standing add <schedule> <timezone> <description> then /tag standing confirm
  • /tag standing list|cancel <id>|pause <id>|enable <id>
Context scope Meaning
focused_reply Explicit reply -> narrow to that parent as evidence.
thread Real thread without explicit reply -> narrow to that thread as evidence.
deictic_recent "上面那张图" / "the image above" -> nearest recent media.
plain Bounded recent text.
Memory tier Behavior
Tier-0 Short-term per-chat buffer, TTL/count evicted.
Tier-1 Long-term, @-derived, consolidated, tombstoned on delete.

Project Structure

  • src/hermes_tag/core.py — config, sqlite store, TagEngine, PlatformSeam.
  • src/hermes_tag/context.py — ContextSelector: the bounded-evidence selector.
  • src/hermes_tag/base.py — TagAdapterMixin: the platform-agnostic orchestration.
  • src/hermes_tag/platforms/feishu.py — Feishu binding: mention detection, media fetch/download, registration.
  • src/hermes_tag/platforms/slack.py — Slack binding: mention detection, media buffering, registration.
  • src/hermes_tag/platforms/dingtalk.py — DingTalk binding: is_in_at_list mention detection, registration (reduced capability — see docs/dingtalk.md).
  • src/hermes_tag/i18n.py — locale strings.
  • src/hermes_tag/adapter.py — back-compat re-export shim.

The platform-agnostic base plus a narrow seam means each platform is a thin add; the generic policy is written once for all of them.

Contributing · Changelog · License