Release Notes: v0.25 – v0.28

v0.25 — GitHub App Webhook

humux can now receive GitHub webhook events directly. When someone opens a PR, pushes a commit, or creates an issue, the agent gets notified in real time. This replaces polling-based GitHub integrations and lets agents react to repository events as they happen.

• Inbound webhook route — receives events from a GitHub App installation, validates the signature, and routes the payload to the configured agent.
• Author login gate — filter events by GitHub username. Useful for ignoring bot-generated events or restricting notifications to specific contributors.
• Stream-capped bodies — large webhook payloads are truncated to prevent blowing out the agent's context window.
• Retryable deduplication — GitHub retries failed deliveries; the webhook handler deduplicates by delivery ID so agents don't process the same event twice.

agents:
  - slug: forge
    github:
      webhook_secret: ${vault:GH_WEBHOOK_SECRET}
      author_gate:
        - mattmezza       # only my events
      ping_chat: -100123  # notify this Telegram chat

v0.26 — Webhook Mention Gate

Building on v0.25's webhook foundation, this release adds a mention gate: agents only process webhook events where they're specifically @mentioned in the body. This prevents noise from high-traffic repositories where the agent should only respond when explicitly called.

• Mention gate — scan the event body for the agent's GitHub login or configured aliases. No mention = no agent turn.
• Configurable ping chat — webhook events trigger a notification in a designated Telegram chat, with a link to the PR or issue.
• Human labels for chats — the agent editor now shows human-readable labels for chat IDs instead of raw Telegram numbers.

v0.27 — Per-Agent LLM Override

Different agents have different needs. A coding agent benefits from the strongest reasoning model. A triage agent that classifies emails can use a cheaper, faster model. v0.27 lets you set the LLM provider and model per agent, overriding the system default.

agents:
  - slug: forge
    llm:
      provider: anthropic
      model: claude-sonnet-4-20250514

  - slug: triage
    llm:
      provider: ollama
      model: llama3.1:8b

• Per-agent LLM config — set provider and model in the agent definition. Subagents inherit their parent's LLM config unless overridden.
• Editor UI tab — configure the LLM override visually in the admin agent editor, with a badge on the agent list showing non-default models.
• Runtime resolution — the turn resolver checks agent-level config first, then falls back to the system default. No code changes needed in tools or skills.

v0.28 — Token Dashboard + CoT Feedback

The largest release in this batch. Two major features: a token usage dashboard for cost visibility, and a chain-of-thought feedback mode for debugging agent reasoning.

Token usage dashboard

• Per-period cards — see token usage broken down by day, week, and month with human-readable units (k, M, B).
• Per-agent attribution — every token is attributed to the agent that consumed it, including subagent tokens rolled up to the parent.
• Top-agent card — see which agent is consuming the most tokens at a glance. Unattributed tokens (from system overhead) are excluded from the ranking.

Chain-of-thought feedback

When enabled per agent, the chain-of-thought mode gives you real-time visibility into what the agent is thinking. Instead of waiting for the final response, you see a thinking indicator as the model reasons through the problem.

• Per-agent toggle — enable CoT feedback mode on agents where you want to see the reasoning process (e.g. the coding agent) and leave it off on agents where you just want the answer (e.g. the triage agent).
• Non-destructive reactions — the thinking indicator uses a Telegram reaction that doesn't clear any agent-set reactions on the triggering message.

Other changes in v0.28

• /subagents command — renamed from /jobs for clarity. Lists running background tasks spawned by the current agent.
• On-demand skill loading — dropped the /zz_skill_* Telegram commands in favor of loading skills on demand when the agent needs them.
• Reminder context folding — delivered reminders are folded back into the origin chat's context, so the agent remembers what it scheduled and why.

Upgrading

All four releases are backwards-compatible. Pull the latest image and restart:

docker compose pull && docker compose up -d

The new features are opt-in: GitHub webhooks require configuring a GitHub App, per-agent LLM overrides default to the system model, and CoT feedback mode is off by default. The token dashboard is always available in the admin UI.