# schmitfe/compute-relay

Source: https://glimind.com/server/github%2Fschmitfe%2Fcompute-relay
Registry id: `github/schmitfe/compute-relay`
Status: unknown

## Tools (34)

- `add_investigation_jobs`: Attach up to 32 existing project jobs. New jobs can instead use prepare_job.investigation_id; revisions inherit grouping.
- `advance_workflow`: Recheck and materialize at most two ready nodes; existing attempts are reused.
- `create_environment_build`: Create a pinned environment recipe; installation runs only later as an approved/budgeted compute job. Only venv uploaded wheels are implemented.
- `create_investigation`: Record a finite scientific plan. Plan limits guide agents; the project grant enforces spending. Executes nothing.
- `finish_investigation`: Close scientific scope and record a report. Optionally cancel drafts/pending reviews; admitted work and accounting remain visible. Export verified results with 
- `get_broker_capabilities`: Optional: inspect enabled extensions when an extension tool is needed. Ordinary job planning uses get_project_context.
- `get_compute_guide`: Read shared guidance on demand; job_request includes typed resources and a small valid example.
- `get_environment_build`: Read build/provenance, observed runtime inventory and the published environment reference. permitted_for_jobs reflects the project's automatic-use policy or exp
- `get_environment_build_policy`: Read explicitly enabled builders, bootstrap environments, resource/storage limits and target platform.
- `get_environment_inventory`: Read the actual observed Python version, installed packages and fingerprint, with observation time and managed-revision drift status. A registered name/version 
- `get_investigation`: Read grouped jobs, abandoned drafts, aggregate accounting and closeout blockers. Wait at most 20 seconds for a revision change.
- `get_job`: Read status, execution progress and completion blockers. detail=True includes package/history. after_revision waits up to 20 seconds; unchanged timeouts return 
- `get_job_artifact`: Read one authorized artifact as bounded base64, without executing or rendering it.
- `get_job_completion`: Read a paginated run manifest: provenance, submission, outputs, accounting and delivery evidence. Reading changes no execution state.
- `get_job_logs`: Read a bounded live-log window. Reuse next_cursor; eof drains the snapshot, finished marks terminal execution, and truncated/reset flag gaps.
- `get_project_allowance`: Refresh balances, reservations, debt, expiry and admission counts. detail=True includes full allowance metadata.
- `get_project_context`: Read current project authority, targets, environments, allowance and recovery inbox; no project ID is needed.
- `get_project_storage`: Read approved target IDs, advisory budgets and cached capacity/unknowns. Does not probe or scan storage.
- `get_workflow`: Read workflow nodes, concrete job IDs, blocked inputs and errors.
- `inspect_environment`
- `list_compute_targets`: Optional focused refresh of permitted clusters and effective_limits. get_project_context already contains these targets. Does not log in to a cluster.
- `list_investigations`: List this project's investigation identities and scope state; paginate with next_cursor.
- `list_job_artifacts`: List registered outputs belonging to a project attempt.
- `list_project_jobs`: List recoverable project work or filter by investigation. Follow next_cursor; reading never executes or cancels work.
- `plan_storage`: Dry-run declared peak byte/inode growth against selected targets. Creates no jobs, files, or reservations; does not enforce OS quotas.
- `prepare_environment_build`: Render the immutable build script/package for review without running an installer.
- `prepare_job`: Prepare an immutable package without executing it. Use current target/environment and staged file descriptors; optional investigation_id groups the stage.
- `prepare_workflow`: Seal at most 32 job nodes and success-only artifact edges. Omit project_id in node requests; executes nothing.
- `request_environment_build`: Request normal project approval or budgeted execution; builds consume compute allowance and submission slots.
- `request_environment_packages`: Build a new usable Python environment or update a managed one through a new immutable revision. packages are uploaded wheel descriptors {name,version,filename,s
- `request_execution`: Request execution and reserve budget; returns approval reasons and state. Retries reuse the default job key; a new explicit key means a deliberate repeat.
- `revise_job`: Prepare a linked revision after review feedback. Omit project_id; changes need a new execution request.
- `start_workflow`: Start a bounded run; each ready job needs normal budget admission or human review.
- `upload_job_file`: Upload one base64-encoded source file (512 KiB maximum); path is a logical name.

Measured by Glimind (https://glimind.com): MCP handshake + tool list only, no tool is ever called. CC BY 4.0.
