Changelog

What changed in each released version, and why. An entry is written when a version is released, summarizing everything since the last one; work in progress lives in the git history. Changes that come from the learning loop say which kind of incident or review taught them, in general terms.

The package follows semantic versioning: a new skill or template is a minor version; a change that breaks how an existing skill is used, or where it writes, is a major version.

0.1.0

The first release.

• A router, sre, that points a situation to the right orchestrator.
• Twelve orchestrators, one per moment of the work, each with its template: define-slos, review-design, plan-migration, review-release, plan-maintenance, design-alerts, find-blind-spots, respond-to-alert, contain-security-incident, reconcile-drift, write-postmortem and learn-from-incident.
• Nine disciplines: reframe-first, scope-and-signal, reversible-first, state-is-special, evidence-before-claims, verified-impact, paved-road, idempotency and expand-contract.
• Two calculators: error-budget.py in define-slos and burn-rate.py in design-alerts.
• Docs: the method, the lifecycle, the layers, the landscape, the glossary and the review rules.
• Two evals, drawn from real runs on an over-the-air update fleet.

Shaped before release by real use:

• Two rounds of runs by agents that had never seen the package, on a working prototype, found gaps in the skills themselves: templates without a place for what the steps asked, labels that didn't match, skills that assumed a web service with traffic, and "done" criteria that skipped the steps that found the most. Each became a fix and, where it applied, a review rule.
• A live incident, a release that crashed after its canary check, was handled with respond-to-alert and written up with write-postmortem. The postmortem's lesson became part of respond-to-alert: a fix shipped as a new release is confirmed past the failure's delay.
• An independent review of the whole package corrected security references (session revocation for leaked keys and stolen tokens), Kubernetes details and the calculators' input checks.