Skip to content

Example workflows

Install any example from Mission Control’s Examples page or the CLI. Check what it needs first:

Terminal window
npx azhi example list
npx azhi example needs ai-sdlc-v2
npx azhi example install ai-sdlc-v2 --repo OWNER/REPO --set slack_channel=...
ExampleWhat it doesExecutor
quality-report (the flagship)Weekly report: tools and a Python script compute metrics; an agent explains them with numbered citations; Slack post gated by a CEL guard. Its tools return fixture datamodel agent
ci-digestCI digest to Slack; the smallest end-to-end examplenone (no model)
ai-sdlc-v2GitHub issue → readiness check → plan → build → independent review and verification → approval → draft PROpenCode
pr-reviewFour read-only reviewers in fresh checkouts, a verifier on a different model family, triage by script, one sticky PR commentOpenCode (Copilot models)
issue-investigationEvidence-backed root-cause analysis with a challenger; an approval gates the commentOpenCode
issue-splitter / issue-splitter-jiraSizes a ticket against the code; if too big, splits into ordered, linked sub-issues after approvalagent
doubler, caller, loop-demoTiny fixtures for subworkflows and loopsnone
ci_runs ┐
flaky ├─► metrics (script) ─► analyse (agent + dataset) ─► report ─► post (Slack, guarded)
incidents┘

Numbers come from the script and are authoritative; the agent explains them. Every claim cites an immutable excerpt, with per-source as-of times.

pr ─┬─ correctness ─┐
├─ security ────┤
├─ tests ───────┼─ verify ─ triage ─ summarize ─ report ─┬─ post (PR comment)
└─ quality ─────┘ └─ notify (Slack)
  • Each reviewer works in its own fresh checkout of refs/pull/<n>/head, with read-only tools only (read, grep, glob, skill), an isolated git and HOME, and a read-only token that is never written.
  • verify (a different model family) tries to refute every finding; triage keeps those confirmed at 80+.
  • A closed or merged PR skips every later step unless allow_closed is true.
  • The comment is ledgered, deduplicated, guarded to the PR under review, and sticky: a later run edits it.
  • Runs on GitHub Copilot models: azhi copilot login.

issue → investigate → challenge → calibrate → report → (approve → post). The RCA holds severity, complexity and confidence with reasons, origin (regression, long-standing, original, unknown), a 5-whys chain with path:line evidence, and a proposed fix with tests to add.

Rules, not the agent alone, decide single or split: a split needs 2+ proposed slices and any of more than one deliverable, more than split_max_files (5) files, or a sensitive area in more than 2 files. A person approves and may leave slices out. Something that is not code stops, and nothing is created. Run it before the delivery workflow.