Published installer available. Native workflow diagrams are a preview; check your installed release’s capabilities.

How it works

One task. The whole picture.

Agree on the plan. Inspect the code and evidence. Decide how to respond to review. Follow the whole task without losing its context.

01 Task journey 02 Evidence 03 Lifecycle 04 Architecture

01 / Task journey

An agreed plan. Reviewable work.

The native workflow starts with your explicit request and keeps plan approval, evidence and PR feedback connected to that task.

Request → approve a plan → coding → checks and separate AI reviews → draft PR → review feedback → approve selected fixes → fresh checks → your team’s merge decision.
Quick asks for one plan approval. Full adds requirements and design approvals. Review feedback returns through an owner decision, then fresh evidence. Open full size

Before coding

Read the plan in Slack Canvas or your coding tool. Approve that version or request changes. Required project checks stay on.

Before PR creation

After the selected checks pass, separate code and security reviews start automatically. Native workflow readiness requires matching results; missing or stale evidence stays visible.

During team review

Comments become proposed fixes. The owner decides which changes to authorize. Original comments and merge state stay in GitHub.

Follow a Quick task →

02 / Code and evidence

The results belong to the code.

Inspect named check output and separately attributed AI reviews. If the candidate changes, results for the earlier code do not establish readiness for the new code.

Candidate A has checks and separate code/security AI reviews. A code change creates candidate B and makes dependent evidence stale. Candidate B needs fresh checks and reviews; failures and missing evidence block native workflow readiness.
The native workflow binds required check and review results to the candidate. AI review findings support inspection; they do not prove correctness. Open full size

Tests and checks

Inspect the command, environment, output and candidate reference. Missing output is a gap, not a passing result.

Separate AI reviews

Code and security reviewers operate in read-only sessions. Their findings are attributed to the reviewer and model.

Your repository’s CI

GitHub CI and branch rules remain authoritative for your repository. Native workflow readiness and repository merge rules are distinct.

Existing direct tasks and historical benchmarks

The existing direct-task path can return draft PRs with failed or unverified checks described. Its publication behavior differs from native workflow readiness. Historical benchmark graders are another evidence layer; they do not qualify a current workflow candidate.

Understand the evidence layers →

03 / Workspace and task lifecycle

Come back with the context intact.

Waiting for a decision is not a failed task. Continuing work is not starting over. Workspace lifetime and the retained task record are separate concerns.

Start and plan → wait for owner → coding and checks → PR and feedback loop → wait for all required PRs → preserve the record → clean up Canvas copies. The workspace can retain context during follow-ups; storage costs remain.
Task artifacts and decisions live in Rovara Code under configured retention. At closeout, verify their preservation before deleting task Canvas copies. Slack replies and GitHub comments remain in their services. Open full size

Wait and continue

Inspect the next responsible action. Idle compute can stop while files remain; retained storage still costs money.

Preserve the record

The task record retains artifact references, decisions and evidence according to configuration. It is not an indefinite-retention promise.

Close deliberately

Cancel work, close a workspace and remove a deployment are different actions. Preserve unpublished changes before closure.

Workspace actions and costs →

04 / Architecture

Know what runs where.

Slack and coding tools are entry points. Rovara Code coordinates the task in your AWS account. GitHub owns PR state; configured providers and connectors have their own data boundaries.

Outside AWS: Slack, coding-tool MCP, authenticated browser client, GitHub and model/connector services. Inside your AWS account: authenticated control plane, workflow record and decisions, EC2 workspace with coding, check execution and separate read-only review sessions, DynamoDB state, S3 artifacts and encrypted EBS files.
Role-level architecture. Separate review sessions do not imply separate machines. Authenticated browser requests and signed GitHub events pass through the control plane. Open full size

Coordinate

The authorized control plane records tasks, decisions and artifact references. GitHub events are checked against the task’s registered PRs.

Work and inspect

The coding worker uses the configured project environment. Checks execute separately from the coding agent’s success claims; reviewers use separate read-only sessions.

Keep ownership clear

The account is the documented security boundary. External model providers and vendor MCP servers receive their configured requests outside that boundary.

Engineering view: routing, state and authority
Slack
Signed ingress → request queue → Fargate orchestrator → control-plane API. Plans use Canvas; concise updates and decisions stay in the task thread.
Coding tools
Local agentx mcp → authorized developer task API. The coding client is an entry surface; the remote harness implements the task.
Worker operations
Control-plane dispatch → operation queue → EC2 worker. Coding and read-only review operations have different tool permissions.
State and files
DynamoDB stores task/workflow/operation state; S3 holds artifacts and session data; encrypted EBS holds workspace files.
PR feedback page
An authenticated browser view reads task-scoped review details and submits owner decisions through the API. A link alone does not grant access.
GitHub
Signed events are deduplicated and scoped to linked PRs. GitHub API state remains authoritative; human merge decisions remain in the repository.

Environments in one AWS account are not independent security boundaries. This role diagram is not a complete network or security assessment.

Architecture reference →

The final decision

You decide what gets merged.

Your team reviews the candidate, task evidence and repository CI with its existing merge controls.

Follow a first task →