Changelog

Custom instructions are added to every run: chat, alert investigations, and pull request replies. You can turn built-in skills on or off and write your own, such as a rollback runbook, which Corvo reads when a task calls for it. MCP servers connect from the same page, with OAuth for servers that need sign-in. When a server's authorization expires, its row shows an Authorize button.

Corvo writes a daily summary of errors, investigations, and pull requests. Set your timezone and a push time in Settings and the report arrives in Telegram, or send /daily_report to the chat bot in a direct message. Each report also has its own page.

You can sign in with a one-time code sent to your email address. Google and GitHub sign-ins that use the same email now open the same account.

Integrations, agent settings, alerts, and reports now belong to a workspace. Create one for each team and switch between them from the sidebar. Invite teammates by email as an Admin, Editor, or Viewer; a setting your role can't change is locked and says who can. Existing accounts were moved into a workspace with their data, and admins can now disconnect GCP or the GitHub App and delete a workspace.

Corvo collects alerts for five minutes, then an orchestrator hands each problem to an investigation. A repeat of a known error goes back to the investigation that already handled it, and a multi-line stack trace now counts as one alert. Each investigation page shows its status, its alerts grouped by error, the agent's summary, and a Cloud Logging link for the whole time span. The agent works out which repository a log came from, so the service-to-repository mapping is gone.

You can message the agent while it works. Queue a follow-up, press Send now to change its direction, or stop the run; stopping keeps your queued messages so you can still send them. Runs no longer stall after 60 seconds, and the conversation stays in sync across tabs.

When the agent learns something about your systems, it proposes a fact on the Memory page. A member approves the fact before the agent uses it, and members can add, edit, or remove facts themselves. Admins can turn on auto-approve after a warning that explains what it allows.

Add the repositories Corvo works in and a setup command for each. Corvo runs every command first, then builds a sandbox snapshot, so each run starts with the code and its dependencies installed.

Alert streaming now reads GKE and Compute Engine logs as well as Cloud Run. Corvo lists the clusters, namespaces, containers, and instances in your project and builds the log filter from what you select. Selecting a whole platform also covers anything created there later, and the current sink keeps running until the new filter is in place. The GCP setup wizard can go back a step, and each step offers a manual path and a script.

Telegram now has two bots: one you chat with, and an alerts bot that only posts incidents. Add the alerts bot to a group and each alert updates in place while its investigation runs.

Corvo now works in Discord. Pick a server and a channel for alerts, and each alert posts there and updates as the agent works. Mention Corvo in the channel to ask about one.

Install the Corvo GitHub App instead of pasting a personal access token. Pull requests and comments now appear under Corvo's bot identity. Corvo also reads repository guidance files such as CLAUDE.md and AGENTS.md, then follows the repository's pull request template when it opens a fix. The personal token connection flow has been removed.

Comment on a pull request opened by Corvo to ask about the fix or request a change. Corvo replies in the same GitHub thread with the current diff and pull request context. Reviewers can follow the discussion without moving to Telegram.

Corvo now works in Telegram group chats. Bind a group to a project by project name or chat ID from Settings. The group receives the same alert, investigation, and pull request updates in one thread.

An alert now reaches every Corvo user connected to the same GCP project. Telegram posts later investigation updates as replies to the original alert, keeping each incident in one thread.

The new read_metrics tool lets Corvo inspect Cloud Monitoring data while it investigates an alert. GCP errors are presented as readable messages instead of raw JSON. The Incidents page also shows the deduplication decision and cached pull request status.

Corvo classifies every incoming alert as drop, skip, or process. A duplicate of an active investigation is dropped, while a repeated alert with an existing fix can be skipped. Alerts that still need work start a new agent run.

When Corvo processes a recurring issue, it includes context from the earlier investigation and fix. The next run can see what was already tried instead of starting with an empty history.

Enter a GCP project ID and choose which supported services Corvo may read. The page lists the APIs, service account, and read-only roles that Corvo will configure before Google sign-in. Corvo then runs the setup steps and verifies the connection.

The agent runtime now handles exhausted action limits and empty model completions. It retries when recovery is safe and returns a clear error when it cannot continue. Sandbox setup is also skipped when a run does not need it.

The Agent Guardrail checks each git and gh command against the permissions set for the user. It also parses compound shell expressions, so a blocked operation cannot be hidden after another command. The Permissions page controls individual actions such as commit, push, merge, release, and workflow reruns. Actions that can publish or discard work require explicit trust.

The app streams the current investigation status and each completed step. Disconnecting closes the live view without stopping the agent. You can return later to see the result.

Alert ingest can now be enabled and dispatched per user. This supports gradual rollout without changing routing for everyone connected to a project.

The public site now includes terms of service and privacy pages. Both pages are linked from the product for easier access.

The new edit_file tool replaces a specific string in place. Corvo no longer needs to rewrite a whole file for a small change. Pull request diffs stay focused on the lines that changed.

Every alert conversation now runs in an isolated sandbox. Concurrent investigations use separate repository clones and working state. A change in one investigation cannot leak into another.

Corvo checks for duplicate alerts during webhook ingest. If the same alert arrives while an investigation is active, the duplicate does not start another agent run. This prevents competing fixes for one incident.

GCP onboarding now generates the required log sink setup script. The page guides you through running it in Cloud Shell and returning to Corvo when setup is complete.

Every alert Corvo triages is now collected on a brand-new Incidents page. From there you can browse the full history of triggered alerts, jump straight to the matching error in the GCP Cloud Logging explorer, follow the GitHub PR whenever Corvo opens a fix, and watch the agent's status in real time as it moves from triage to investigation to action.

Corvo's replies in Telegram are no longer one long block of plain text. The new Telegram notifier supports bold text and clearer paragraph structure, so RCAs, investigation results, and fix suggestions are much easier to read.

We made a number of performance and stability improvements to the runtime. Long-running tasks now move to a background process automatically, and alert triage no longer mixes in unrelated recent chat history. The result is more accurate PR fixes and lower latency overall.

Corvo used to send a notification for almost every intermediate step during triage. From v0.0.3 the flow is condensed to three key moments: the initial notification, the RCA once investigation completes, and the final investigation result. You get far less noise while keeping full incident coverage.

New users now see a language picker during onboarding to set which language the Corvo agent replies in, and you can change it anytime under Settings > Basic. English and Traditional Chinese are available today. The setting flows through every Corvo workflow, so Corvo speaks your chosen language everywhere it has something to say.

The first end-to-end product surface: a public landing page, the authenticated app shell with sign-in, and a GCP onboarding flow that guides users through wiring Cloud Logging alert notifications and a least-privilege impersonation service account.

Cloud Logging error notifications hit a signed webhook that verifies the HMAC, normalizes the payload, and hands the alert to the SRE alert workflow. The workflow runs a structured triage → investigate → act sequence on the orchestrator agent, with per-step prompts and a strict turn protocol.

The agent server runs on Mastra with a sandboxed workspace (E2B or local) where the agent can clone the affected repo, read code, run bash commands, and propose fixes. Service-to-repo mappings let the agent figure out which repository owns the alerting service, and a UI for managing those mappings ships in the same release.

Corvo notifies users in Telegram as the agent works — bind a chat once via the settings flow, and the agent will keep the channel updated through the lifecycle of every incident.

Cloud Run deploys for both the web app and the agent server, env vars sourced from GCP Secret Manager via pull-env.sh and push-env.sh, the agent runs as a least-privilege service account (corvo-agent-runtime), and a separate Cloud Tasks queue handles asynchronous agent dispatch. A hotfix at v0.0.2 wires the prod workflow to the right Cloud Tasks queue.