UNPKG

@kontextmind/kxm

Version:

KXM local-first multi-agent orchestration and operator dashboard

48 lines (34 loc) • 1.98 kB
--- name: kxm-triage description: KontextMind knowledge plane only, the separate kontext CLI and km_ tools, not KXM or this plugin's kxm_* tools. Works the KontextMind review queue with km_review. Use only when the user names KontextMind, the kontext CLI, or a km_ tool. license: Apache-2.0 compatibility: Any agent that can run a shell or MCP client. No vendor-only tools. metadata: workflow: memory-triage version: "0.1.1" argument-hint: "[list|promote|skip|research|suspicious]" complete: "km_review list, km_review resolve, kontext review list" suite: kxm --- # kxm-triage Beacon — `km_status` with `skill: "kxm-triage"`. ## Queue 1. `km_review` `action=list` (optional kind filter). Group by kind — learning, drift, contradiction, gap, loop, suspicious. 2. **Suspicious first.** Secret-gate trips resolve before anything else. Never paste the matched secret; name the rule only. 3. Resolve with `km_review` `action=resolve`, `id`, `verdict`, `reason`. - `promote` — true, correctly scoped, deduped. Server commits to curated pages. No PR ceremony. - `research` — needs evidence; say what is missing. - `skip` — noise; mandatory reason (duplicate, wrong-scope, low-value, false-positive). - `suspicious` — gate/threat follow-up. CLI — `kontext review list [--kind]`, `kontext review resolve <id> <verdict>`. ## Kind notes - Drift — verify against repo HEAD. `doc_claims` is the probe evidence. - Loop/gap promotions become artifacts. Insight `promoted_to` must point at the resulting page or skill. - Three dismissals of the same insight type mute it 30 days. Do not dismiss unread items. ## Rules - Every verdict has a reason. - Strict namespaces — show author, source session, SHA to the human before promote. ## Autocomplete Slash hint — `[list|promote|skip|research|suspicious]` Complete — `km_review list, km_review resolve, kontext review list` Works on any harness. Catalog — `plugins/kxm/skills/hints.json`.