Aruing

English | 中文

A Kubernetes ops agent that answers with tools and evidence—not vibes.

Ask in natural language. The agent reasons, calls cluster tools for real evidence, forms hypotheses, and—when you need root cause—runs a forced evidence-adjudication chain so every conclusion is traceable.

Tools go through a shared Registry / Dispatcher (shell-less kubectl backend, policy for auth). Model output never pretends to be Evidence.

Version 0.1.0 — askable diagnostic assistant (in progress). Milestone 0.1.0-beta7 (post-compaction range rehydrate) closed; 0.1.0-beta6 (deep follow-up by run) closed.

• aruing run — single-shot diagnosis via linear Orchestrator
• aruing chat — multi-turn Session + Tower (LLM required); escalate when root cause is needed; successful escalate writes RunLedger; follow-ups inject prior_run_details (explain via reply by default); after compact, range rehydrate restores Store text into rehydrated_messages when needed

Details: docs/project-state.md.

Run → Query → Target → Hypothesis → Task → Evidence → Verdict → Report

• Run — one diagnosis unit
• Query / Node / Edge — unverified clues from the question
• Target — objects confirmed in the real cluster
• Hypothesis — candidate causes awaiting evidence
• Evidence — records from actual tool execution (only trusted fact source)
• Verdict — only from Evidence
• Report — cites Verdict + Evidence; does not invent

make build              # build cmd/aruing
make test               # all tests
make check              # full CI (test-ci + vet + lint + fmt + tidy + vuln)

./bin/aruing run why is demo-api in default unreachable
./bin/aruing run --format json why is demo-api in default unreachable
./bin/aruing chat hello                                    # multi-turn (LLM required; session id on stderr)
./bin/aruing chat --session sess_xxx check redis again

Repo ignores .env. Copy the template; Make loads it:

cp .env.example .env    # BaseURL / APIKey / Model
make print-env
make run-llm
make run-llm QUESTION='why is demo-api in default unreachable'
# make run-llm ENV_FILE=.env.ollama

All three LLM vars set → real agent roles; any missing → fakes (same as no .env).
aruing chat / make chat require all three; no silent fake path.

make chat
make chat CHAT_MSG='hello'

• Flat entities linked by RunID—no nested Run
• Clues are not Targets until the environment confirms them
• Model output ≠ Evidence; Verdicts must cite Evidence
• No enumerating user ops / resource types; no artificial N-item amputations of normal capability (over budget → compact, don’t silently drop)
• Tools are not inherently R/O; policy gates execution. Read tools registered now; write tools later with approval
• run → Orchestrator; chat → Session.Turn + Tower; same Dispatcher

Full list: docs/architecture.md (incl. #15–#18).

Longer design notes live in a private arui-note/aruing/ notebook (maintainer only).