LLM-powered issue triage. Built for builders.

LLM-powered triage agent for GitHub issues and PRs. Part of the clagentic suite.

What it does

clagentic:triage watches a GitHub org or repo for new issues and pull requests. By default only external contributors are triaged — issues and PRs from the operator's own org members, owners, and collaborators are filtered out, along with any logins the operator chooses to ignore. This is configurable via source.watch_associations, source.ignore_logins, and source.watch_logins (see docs/CONFIG.md). Each inbound event is enriched with repo context (a per-repo intent file plus referenced documentation), then assessed by an LLM against the stated intent — does this issue belong here, does this PR meet the bar, what action should follow?

The LLM returns a verdict and a suggested action. By default that suggestion waits in a human-review queue. A human approves, overrides, or rejects it via the CLI. Auto-approve can be enabled per action class once you trust the model's verdicts on that class.

Approved actions execute through the source adapter (post comment, request changes, close, approve PR) and optionally dispatch into a backend task system (Jira, Linear, GitHub Issues, a webhook endpoint, or any custom dispatcher).

Key properties

• Human-in-the-loop by default. Every verdict waits for approval unless you explicitly opt an action class into auto-approve. Confidence below the configured threshold always routes to HITL regardless of auto-approve settings.
• Multi-runner LLM. Four backends: claude-cli (default — spawns the claude CLI, requires OAuth session on the same host), anthropic-api (direct HTTP, no SDK, just an API key), openai-compatible (covers OpenAI, Azure OpenAI, Ollama, clagentic:router, and any OpenAI-compatible server — recommended for containerized or multi-host deployments), clagentic-router (legacy private protocol; prefer openai-compatible against a router instance). Model selection is config-driven, never hardcoded.
• Pluggable source adapters. Adapters normalize events from a platform (GitHub, GitLab, Forgejo, ...) into a common Event schema. The pipeline is adapter-agnostic.
• Pluggable dispatch backends. Dispatchers push verdicts into backend systems. Any backend — including private or internal tools — plugs in by module path. The core ships generic reference dispatchers (webhook, noop); concrete backends are external packages.
• Post-verdict hooks. Hooks fire after assessment, before the queue. The bundled clagentic-console hook opens a clagentic:console conversation on escalate verdicts. Additional hooks load by module path.
• Zero runtime npm dependencies. ESM, Node 20+, no bundler required.

Install

Requires Node 20 or later.

Run from a clone:

git clone https://github.com/clagentic/clagentic-triage
cd clagentic-triage
node src/cli.js run

The binary is clagentic-triage.

Quickstart

Create a minimal config file at ~/.config/clagentic/triage/config.json:

{
  "source": {
    "adapter": "github",
    "org": "your-org"
  },
  "runner": "claude-cli",
  "model": "claude-sonnet-4-5"
}

Set the required token:

export CLAGENTIC_TRIAGE_GITHUB_TOKEN=ghp_...

Run a single triage pass:

clagentic-triage run

Run a continuous poll loop:

clagentic-triage watch

Review the pending queue:

clagentic-triage review

Approve, override, or reject a queued verdict:

clagentic-triage approve <id>
clagentic-triage override <id> --action <class>
clagentic-triage reject <id>

override approves the item but substitutes a different action class (respond, request_changes, close, dispatch, escalate, approve) in place of the one the LLM suggested.

Configuration

Configuration is loaded from environment variables (CLAGENTIC_TRIAGE_* prefix), then triage.config.json in the working directory, then ~/.config/clagentic/triage/config.json.

See docs/CONFIG.md for the full schema and all environment variable overrides.

The two-tier pre-filter is optional and off by default. When enabled, a fast cheap LLM pass classifies each inbound event as noise or real before the main assessor runs, reducing cost on high-volume repos. Enable it by setting pre_filter.enabled: true in the config file once you have a representative event sample to tune against. See docs/CONFIG.md#pre_filter and DD-011 in docs/DESIGN-DECISIONS.md.

Security model

• Webhook verification. Inbound webhook deliveries are verified by the source adapter (HMAC-SHA256 for GitHub, token comparison for GitLab/Forgejo) before any payload parsing. The webhook secret is required at startup when the webhook server is enabled — an empty secret is rejected at config load time.
• Prompt-injection boundary. Issue and PR bodies are wrapped in <UNTRUSTED_USER_CONTENT> tags and pass through a redaction step before LLM prompt construction. Common secret patterns (ghp_*, sk-*, AKIA*, -----BEGIN *) are redacted to [REDACTED].
• Intent file trust boundary. The per-repo intent file (.github/triage-intent.yml) is authored by repo maintainers with write access. Its content is injected into the prompt as operator-controlled configuration, not user input. Referenced context files are path-validated against an extension allowlist before fetching. Intent files are capped at 64 KB; referenced context files are capped per-file and in aggregate.
• Secrets via env, never stored. Tokens and API keys are read from environment variables named by runner_api_key_env (or the runner's default env var). They are never written to the config file or the pending queue.

See docs/DESIGN-DECISIONS.md for the full rationale behind each security decision (DD-001 through DD-011).

Documentation

Document	Contents
docs/ARCHITECTURE.md	Pipeline stages, module layout, interfaces
docs/CONFIG.md	Full config schema and environment variable reference
docs/GITHUB_APP.md	GitHub integration: PAT vs GitHub App, setup, webhook ingress, reverse proxy
docs/ADAPTERS.md	Source adapter interface (poll + webhook), how to write one
docs/DISPATCHERS.md	Dispatch backend interface, reference dispatchers, third-party plugins
docs/SECURITY.md	Token security, input validation, PII and data residency
docs/DESIGN-DECISIONS.md	DD-001..DD-011 rationale and security decisions
docs/INTENT_AUTHORING.md	How to write a .github/triage-intent.yml file for your repo

Support

If clagentic:triage is useful to you: ko-fi.com/clagentic

Disclaimer

Not affiliated with Anthropic or OpenAI. Claude is a trademark of Anthropic. Codex is a trademark of OpenAI. Provided "as is" without warranty. Users are responsible for complying with their AI provider's terms of service.

License

FSL-1.1-MIT — Functional Source License 1.1, with MIT as the Change License.

Free for personal, internal-business, evaluation, research, and non-commercial use. Not free for offering this tool (or a substantial fork) as a competing commercial product. Each release auto-converts to MIT on its second anniversary.