Skip to content
Aftershock

Changelog

Every release, newest first. RSS feed.

  1. v0.2.0

    Incident memory goes beyond Python and TypeScript, reaches the tools teams already use, and proves it is still working. The website and app are redesigned.

    Proof

    • Still guarding: weekly, each fix is undone on today's default branch and its test must fail again; tests are marked Weak or Drifted when they no longer would (Labs).
    • Test strength: mutation testing of the fixed lines measures how precisely a test pins down its fix (Labs).
    • Several proven tests per incident, each proven on its own, for example a unit test and an API test.
    • Languages: Go, Java and Kotlin (JUnit with Maven or Gradle), Ruby (RSpec) and Playwright browser tests, each proven in Docker (Labs). Monorepo packages via packages: in .aftershock/config.yaml.
    • Install cache: trusted runs snapshot installed dependencies so later runs skip the install (Labs).

    Guard

    • Recurrence risk on every pull request check, even when no test runs.
    • Merge queue checks on the merge-group commit (Labs).
    • Seen it before: new incidents are linked to past ones they resemble; merged pull requests that change the files an open alert points at are suggested as its fix.
    • Flaky tests are quarantined after repeated failures and released after seven green nights.

    Integrations

    • incident.io, Linear, Rootly, Opsgenie, Datadog, Jira and a generic Standard Webhooks endpoint, alongside Sentry and PagerDuty (Labs). Replay window, size caps and constant-time verification for every source.
    • Notifications to Slack and signed outbound webhooks, with retries, auto-disable and a weekly digest (Labs).
    • GitHub Action MaanavKrishna/Aftershock/action runs checks on your own runners.
    • MCP server for Claude Code, Cursor and Copilot (check_change, ask, lessons and incidents), and lessons synced into AGENTS.md by pull request (Labs).
    • README badge for public repositories (Labs).

    Insight and evidence

    • Insights: proof rate, time to proof, recurrences blocked, test health and strength, fault lines and an 8-week trend, with a shareable read-only reliability report.
    • Ask Aftershock: questions answered only from your incidents, always with citations.
    • Signed audit evidence: Ed25519 over RFC 8785 JSON, with a public JWKS, for SOC 2 and DORA.
    • Backfill: import up to 25 closed incident issues at once.

    Platform and trust

    • Read and write API token scopes; audit log; deletion of incidents, repositories, workspaces and accounts; 30-day retention; uninstall purge.
    • Plans and active-committer metering; every run type counts toward sandbox hours; automatic fallback to GitHub Actions at the limit.
    • Structured logs, feature flags, an operator kill switch, a public threat model and operations runbook.
    • Self-hosting with Docker Compose.

    Design

    • New design system (DS 2) with light and dark themes, and a rebuilt website: proof replay, product tour, agents, evidence, integrations, compare, pricing, security, changelog, and multi-page docs with search. Accessibility checked with axe in both themes.
    • App redesign: dark mode, ⌘K command palette, activity inbox, mobile navigation, animated proof cells and a searchable log viewer.

  2. v0.1.0

    First public release. Live at https://aftershock-mk.vercel.app.

    Incidents

    • Capture from labelled GitHub issues (Trigger / Observed / Expected sections are split out), pasted postmortems (the model extracts the fields and the repository), a form, or signed Sentry and PagerDuty webhooks.
    • Incidents without a fix wait in Awaiting fix; merging a pull request that mentions INC-n, or closing the issue with a pull request, starts time travel.

    Time travel

    • A model drafts one test in the repository's framework (pytest, vitest or jest) and style.
    • The test is admitted only if it fails 3/3 on the commit before the fix and passes 3/3 on the fix, in an isolated Vercel Sandbox microVM (Docker locally). Rejected drafts are redrafted with the failure as feedback.
    • Proven tests are proposed to the repository in a bot pull request with a .aftershock/incidents/INC-n.yaml record.

    Guard

    • Every pull request runs the relevant memory tests on GitHub's test merge commit and gets one check run and one comment: Recur, Safe, Inconclusive or Skipped.
    • Suggested fixes are posted only after passing every memory test for the repository.
    • Retro-check runs a newly proven test against every open pull request.
    • Epicenter bisects history to find the commit and pull request that introduced the bug.
    • Nightly re-checks track test health (Healthy, Flaky, Failing on main) and export LESSONS.md for coding agents.

    Platform

    • One GitHub App for sign-in and repository access; organisation members join installed workspaces on sign-in.
    • HTTP API, aftershock CLI, and a GitHub Actions runner mode.
    • Pluggable model: Muse by default, Vercel AI Gateway, or any public OpenAI-compatible endpoint.
    • Next.js 16, Neon Postgres, Vercel Workflow and Vercel Sandbox; runs on Vercel's free tier.