Modal Research Compute

Modal Research Compute

Windows Runtime Commands

On native Windows, use the managed Windows runner and the native runtime command target. Set $runtime to the installed runtime root. Multi-agent installs usually use %LOCALAPPDATA%\ai-agents-skills\runtime. Then run:

$runtime = if ($env:AAS_RUNTIME_ROOT) { $env:AAS_RUNTIME_ROOT } else { "$env:LOCALAPPDATA\ai-agents-skills\runtime" }
& "$runtime\run_skill.ps1" "skills/modal-research-compute/run_modal_research_compute.ps1" <args>

POSIX examples below use run_skill.sh and .sh command targets; use the Windows command target above on native Windows.

Use this skill when the task is about:

• exhaustive search
• object enumeration
• counterexample hunting
• large parameter sweeps
• remote execution of generated experiment code
• GPU-suitable document, embedding, reranking, or tensor workloads

This skill is the integration layer for the local research_compute broker.

When to prefer this skill

• the user wants Modal involved automatically
• the local machine is CPU, memory, disk, or GPU constrained for the requested workload
• the workload is long-running enough that remote execution is a better fit

Unified remote offload

Route through the broker's configured routing_order; the recommended order is local > Kaggle > Modal > Hetzner > GitHub Actions. A valid custom configured order is honored, with local fixed first and supported remote lanes unique, reordered, or omitted. GitHub Actions is the last recommended lane and is appropriate only for proportionate validation of a private repository's own committed experiment code when earlier adequate lanes are unavailable. Follow the contract in references/github-actions-offload.md (driver signature, per-config SIGALRM, no Pool on 2-core runners, merge needs: + non-vacuous banked-value controls, loud-fail semantics, staged pushes, merge-back/resume). The broker checks the GitHub Actions minutes gate before dispatch.

A failed readiness, quota, or budget gate makes that lane unavailable and the router continues through the configured order. Lane checks are lazy, so a safe local decision and an earlier accepted remote lane do not contact later providers. Treat the job as blocked only after the permitted cascade is exhausted, or when an explicit backend override fails its selected lane's gate.

When remote offload is unavailable (minutes exhausted, broker down, unverifiable credit) and computation must run locally, follow references/local-compute-throttle.md (one lockfile-guarded single-process job, idle priority per OS, subprocess-timeout watchdogs, chunked resumable checkpoints, load/CPU guards) — its rules are cross-platform (Linux, macOS, Windows).

Multi-backend parallel fan-out (v2)

For a LARGE divisible batch job — M independent, resumable chunks (a sweep or enumeration split into shards) — the fan-out scheduler splits the chunks across SEVERAL lanes at once (some chunks local, some on the free lane, some on a paid lane), each lane sized to its spare capacity, to minimise makespan while minimising cost. It is a scheduler on top of the same per-lane probes; small jobs keep using the single-lane router. Fan-out is opt-in ([fanout].enabled) and triggers only when the job declares at least [fanout].min_chunks chunks.

Each job carries a policy.speed_cost_weight in [0, 1] (0 cheapest / free lanes only, 1 fastest / recruit paid lanes, 0.5 blend). Every hard rail still binds — per-lane budget caps, the €3/day auto-approve envelope, the GitHub Actions 60% minutes cap, Kaggle's weekly GPU-hour quota, and local's self-preservation load-cap are enforced as per-lane chunk ceilings the knob can never breach. See compute-offload-routing.md for the full contract. Plan a fan-out (no dispatch) with run fanout-plan job.json; a failed lane's unfinished chunks are reassigned to a healthy lane so no resumable work is lost. Completed chunks still consume cumulative lane rails during reassignment, and a successful retry replaces an earlier failed/vacuous duplicate during merge.

Core workflow

Work through compute-offload-sizing-gate first: measure the workload, read the declared lane capacity, and only then size the manifest. The planner sizes the job against modal_backend.FUNCTION_CAPACITY — the same table modal_app.py builds its @app.function(...) decorators from — but it can only enforce the numbers you supply, so an unmeasured peak RSS defeats the check.

1. If local resources matter, run get-available-resources.
2. Build a broker manifest JSON for the task. The resource block must be the top-level constraints key; any other top-level key is rejected (unknown_manifest_key).
3. Run broker plan (or fanout-plan for a large divisible job). Confirm the plan echoes a non-empty constraints block equal to what was submitted, and read Modal's trail reason: modal_capacity_exceeded:ram_gb … means the job does not fit, and cores_oversubscribed=… means it fits but will run at a fraction of the estimated throughput.
4. Use run plan as the decision boundary. Execute a selected Kaggle or Hetzner lane through its corresponding lane skill; run submit dispatches only Modal/GitHub Actions. Execute an accepted local plan under the broker's reported worker limits.
5. Use the selected lane's wait and fetch commands to retrieve results and logs.

Runtime commands

Linux (resolve the launcher for the current agent — a root-owned component generation when one is installed, otherwise the per-user runtime — then call it):

# Credential-bearing lanes must launch from a root-owned AAS component
# generation; the per-user runtime copy is refused by the credential gate.
# The generation directory is named for a git commit, so its name carries no
# ordering -- pick the newest publish, which the store records as its mtime.
launcher="${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh"
newest=0
for gen in /usr/local/libexec/coding-system/components/ai-agents-skills/*/; do
  gen="${gen%/}"
  [ -f "$gen/manifest/credential-runtime.json" ] || continue
  [ -x "$gen/canonical/runtime/runners/run_skill.sh" ] || continue
  stamp="$(stat -c %Y "$gen" 2>/dev/null || stat -f %m "$gen" 2>/dev/null)" || continue
  [ "${stamp:-0}" -gt "$newest" ] || continue
  newest="$stamp"
  launcher="$gen/canonical/runtime/runners/run_skill.sh"
done
run() { bash "$launcher" skills/modal-research-compute/run_modal_research_compute.sh "$@"; }

run bootstrap                          # one-time: generate config if absent, authenticate gh, check deps, doctor
run doctor                             # routing warning + Modal and optional GitHub Actions readiness
run plan        /path/to/job.json
run fanout-plan /path/to/job.json      # v2: split a large divisible job across lanes (no dispatch)
run submit /path/to/job.json
run wait   <job_id>
run fetch  <job_id> --dest /path/to/output

On targets that install a local skill wrapper, that wrapper should forward to the same runtime command target.

skills/modal-research-compute/run_modal_research_compute.sh doctor

Windows:

$runtime = if ($env:AAS_RUNTIME_ROOT) { $env:AAS_RUNTIME_ROOT } else { "$env:LOCALAPPDATA\ai-agents-skills\runtime" }
& "$runtime\run_skill.ps1" `
  "skills/modal-research-compute/run_modal_research_compute.ps1" `
  doctor

Operational notes

• The broker is the decision boundary. Do not call Modal directly from the normal Codex flow when the broker can handle the task.
• The managed broker wrapper consumes AAS_COMPUTE_SECRETS_FILE from the launcher environment. That protected file accepts exactly HCLOUD_TOKEN, HCLOUD_SSH_KEYS, KAGGLE_API_TOKEN, and KAGGLE_CONFIG_DIR; the strict managed loader validates the shared authority without printing values, then removes all four cross-lane keys and the pointer from the Modal child. Modal keeps only its native profile/config or MODAL_TOKEN_ID and MODAL_TOKEN_SECRET authority. Do not put a pointer or credential assignment in a job manifest or loop env.
• CPU-heavy combinatorial workloads should default to remote CPU or high-memory CPU, not GPU.
• GPU use should be explicit in the manifest or clearly justified by the workload.
• doctor and plan work without a deployed Modal app. Modal-backed submission and deploy need a Modal-authenticated host; GitHub Actions wait/fetch use an attempt-unique dispatch id plus the recorded exact GHA run id and gh, not Modal credentials. Unverifiable GHA timing stays reserved, while verified billed-equivalent usage remains accrued through the UTC billing cycle. Kaggle and Hetzner plans hand off to their lane drivers.
• Linux hosts become Modal-ready after python3 -m pip install --user --upgrade modal and modal token set or modal token new.
• Windows hosts should install modal into the selected Python environment and ensure modal.exe is on PATH so broker deploy can find it.
• Broker state persists under the runtime memories tree, while fetched outputs materialize under the caller workspace by default.
• One-time per machine, run bootstrap: it generates research-compute.toml from the example if absent (never overwriting an existing one), authenticates gh, checks deps, and runs doctor. Use this to set up a host that does not have the full system installer.
• GitHub Actions ToS compliance: the broker's gha lane runs only inside a private research repo, executes that repo's own committed experiment code (parameters are data, never executed), is budget-gated, and is the last recommended automatic backend after local, Kaggle, Modal, and Hetzner — never a general compute pool. Configure it under [gha] in research-compute.toml.

Recommended templates

When this skill is involved, consider these workflow templates (install via the workflow-templates artifact profile, or --with-deps to pull backing skills):

• compute-offload-sizing-gate -- Pre-dispatch worksheet: measure the workload, read the declared lane capacity, write the manifest in the correct dialect, assert the plan, and verify the realized allocation.
• autonomous-research-loop-runbook -- Bounded autonomous research-loop runbook with four stop conditions, single-path solving, mandatory cross-agent verification, fresh-agent backtracking, and five-lane broker-routed heavy-compute offload with per-lane safety gates.
• engineering-delivery-loop-runbook -- Bounded build-and-deliver loop runbook: single-path implementation with seen-to-fail proof, cross-agent diff verification, behavior-preserving cleanup, and five-lane broker-routed heavy-compute offload with per-lane safety gates.