AURORA Agent for VS Code
Context engineering, with receipts — compile decision context, drive grant-authorised autonomous
workflows, and manage the AURORA Agent platform from VS Code.
The extension is a front-end for the bioprism command line, the bioprism-mcp MCP server, and
the bioprism-api HTTP gateway. It ships no platform logic of its own: every result comes from the
CLI's own JSON, summaries show limitations verbatim and digests in full, and every summary links to
its raw JSON — the file on disk when the CLI wrote one, otherwise a read-only virtual document
holding the CLI's exact output (dry-run outcomes, verify/explain/validate results).
Research and developer infrastructure: it does not diagnose an individual, recommend treatment,
triage care, enroll participants, or claim medical-device functionality.
What it does
- Compile context — pick a world and a query, run
context compile with --certificate-out
and --section-out into .aurora/, and read the certificate as a faithful summary: verdict,
selection counts, the omission-accounting table, full digests, and the certificate's
limitations array verbatim. Verify and explain commands wrap context verify --certificate
and context explain.
- Workflows view — lists the capability groups from
workflow catalogue --json with their
tools (missing tools stay visible). Context-menu actions scaffold a group (the generated steps
open as an untitled document you review and save) and instantiate it. Instantiate runs the
CLI's authoritative no-dispatch preflight; nothing is executed by either action.
- Autopilot —
grant-template opens an editable grant document (authority for autonomous
dispatch comes only from that document; there is no default grant). Run shows a confirmation
modal listing the grant's allowed_tools, allow_side_effects, and max_attempts before real
dispatch; Dry-Run needs no confirmation because --dry-run dispatches nothing and writes
nothing. Reports land in the reports directory and render with final_status, totals, attempts,
digests, and limitations verbatim; autopilot verify --report checks a report's digest.
- Autopilot Reports view — watches the reports directory (default
<workspace>/.aurora/reports) and badges each report by its final_status
(succeeded / exhausted / refused).
- Pipelines view — if the
gh CLI is installed and authenticated, lists the last 10 GitHub
Actions runs for the repository named by the workspace's remote.origin.url; a click opens the
run in the browser. Without gh the view shows a single informational row.
- Gateway — starts
bioprism-api as a VS Code task bound to 127.0.0.1:8787 (no --token
is supplied; the gateway's own default is loopback), and stops it.
- Doctor — status-bar item plus a command that prints binary resolution,
bioprism --version,
root, MCP API availability, and gh presence to the "AURORA Agent" output channel. That channel
also logs every command line the extension spawns.
Exit code 1 from the CLI is treated as what it is — a verdict, not an error. Failures surface the
CLI's own retryability word (terminal, retryable_after_change, retryable_as_is).
Install
There is no Marketplace listing yet; sideload the .vsix from the release:
- Download
aurora-agent-0.1.3.vsix from the v0.1.3 release
(https://github.com/AURORA-NEURO/aurora-agent/releases/tag/v0.1.3).
code --install-extension aurora-agent-0.1.3.vsix
(or Extensions view → ... → Install from VSIX...).
Build from source (alternative)
From editors/vscode, run npm install && npm run package — it produces the same
aurora-agent-0.1.3.vsix, installed with the same code --install-extension step.
Binary resolution
The extension finds the platform binaries in this order:
auroraAgent.binaryDir setting (a directory containing bioprism(.exe) and friends)
AURORA_AGENT_ROOT environment variable (a checkout with target/release/)
- any workspace folder containing
target/release/bioprism(.exe)
~/aurora-agent, then ~/bioprism
- the managed download cache
If nothing is found, commands offer to download the v0.1.3 release archive for your platform from
GitHub Releases. The download is verified against a SHA-256 pinned in the extension source; on
mismatch the file is deleted and nothing is extracted. Archives exist for Windows x64, macOS
x64/arm64, and Linux x64 (zip extraction uses the system tar).
MCP / AI integration
On VS Code 1.101 or later the extension registers an MCP server definition provider
(contribution point mcpServerDefinitionProviders, API
vscode.lm.registerMcpServerDefinitionProvider), publishing an "AURORA Agent" stdio server that
runs bioprism-mcp --root <root>. MCP-aware AI in VS Code — including Copilot agent mode — can
then call the platform's 259 tools (path-confined to the root). The API is stable in 1.101, which
is why engines.vscode is ^1.101.0; a runtime capability check still guards registration so the
extension degrades gracefully in builds that lack the API.
For other MCP clients, AURORA Agent: Copy MCP Configuration Snippet puts a ready
mcpServers block (absolute binary path plus --root) on the clipboard for pasting into
.mcp.json or equivalent.
Settings
| Setting |
Default |
Meaning |
auroraAgent.binaryDir |
(empty) |
Directory containing the bioprism, bioprism-mcp, bioprism-api binaries; highest-priority source. |
auroraAgent.root |
(empty) |
Data root passed as --root; overrides the detected checkout root. |
auroraAgent.reportsDir |
(empty) |
Reports directory for the Autopilot Reports view and --report-out; empty means <first workspace folder>/.aurora/reports. |
Limitations
- No Marketplace listing yet — install is by sideloaded
.vsix only.
- The managed download contains binaries only: its default root has no fixtures, so
fixture-dependent features (the shipped reference worlds, for instance) need
auroraAgent.root
pointed at a checkout.
- Autopilot has no scheduling or recurrence, no cross-process resume, and no wall-clock deadlines;
a run lives and dies with the process that started it.
- MCP registration depends on the
vscode.lm.registerMcpServerDefinitionProvider API.
engines.vscode (^1.101.0) already blocks installation on older VS Code, so the runtime
guard exists for builds that report 1.101+ but do not implement the API (some VS Code-derived
editors): there the extension's other features work and no MCP server is registered. Doctor
reports which case applies.
- Tests are
node:test unit tests of the extension's pure logic (pinned-hash table, resolution
ordering, CLI envelope parsing, gh run mapping, grant summarising, summary rendering). There
are no VS Code UI tests.
- Research and developer infrastructure: it does not diagnose an individual, recommend treatment,
triage care, enroll participants, or claim medical-device functionality.