Sim CLI Agent Skill

sim-cli

You are working in a sim-cli-enabled solver workflow. Use the sim

runtime where it adds control and observability, and compose solver-native

tools directly where they are the right primitive. This skill carries the

cross-solver discipline — the rules that hold across sim-cli workflows,

regardless of which solver you drive.

The solver's plugin skill is self-contained for solver-specific work:

solver hard constraints, dependency chains, snippets, workflows, SDK/solver

notes. It does not depend on loading this skill — but it should stay

consistent with the discipline below. If a rule applies to more than one

solver, it belongs here, not in a plugin skill.

Execution models

The plugin skill tells you which model the solver uses. Choose the narrowest

correct execution primitive; sim-cli is composable, not a universal wrapper.

| Model | Used by | Lifecycle |

|---|---|---|

| Persistent session | Drivers that hold a live process open | sim connect → sim exec × N → sim inspect → sim disconnect |

| sim-wrapped one-shot batch | Plugin wrappers that add linting, profile-aware invocation, parsing, history, or safer execution | sim check → sim run → parse_output/logs → evaluate |

| Native/tool one-shot batch | Ready solver decks/scripts, vendor batch commands, project pipelines, or post-processing tools where native execution is the right primitive | version/profile probe → native command/script → parse artifacts/logs → evaluate |

sim-cli does not try to wrap every solver operation, executable flag,

vendor batch mode, or post-processing script. Use sim when it adds a stable

control plane: solver discovery and profile checks, live sessions, bounded

exec / inspect loops, history/log capture, parsing, screenshots, or shared

agent discipline. If the artifact is already a solver-native deck/script and

the solver's own batch command is the right execution primitive, call that

native command directly instead of forcing it through sim run. Preserve

stdout/stderr, generated files, and acceptance evidence just as carefully.

Use sim run when the plugin skill says it is the canonical one-shot path or

when the wrapper adds real value. Otherwise, keep the sim-cli discipline while

composing the solver's native tools directly.

The loop, any model: classify inputs and get the missing Category A values

from the user (including the acceptance criterion) → choose the execution

primitive → run a Step-0 version/profile probe (sim inspect session.versions

for persistent sessions; sim check <solver>, plugin guidance, or native

--version / license probe for one-shot work) → execute one bounded step at a

time → inspect last.result or parse the native logs/artifacts between steps →

evaluate against the acceptance criterion → sim disconnect if a persistent

session was opened. On any failure, stop and report — do not silently retry.

Hard constraints (shared sim-cli discipline)

• Never invent Category A defaults. Physical decisions — geometry,

materials, boundary conditions, the acceptance criterion — must come from

the user. "Just use defaults" / "just run it" does not override this;

treat it as a missing input and ask.

• Step-0 version/profile probe is mandatory. After sim connect, call

sim inspect session.versions and use the returned profile /

active_sdk_layer / active_solver_layer to pick the right files in the

plugin skill. Before sim run or native one-shot execution, run the

relevant probe for that path (sim check <solver>, plugin guidance, or the

solver's native --version / license check). If a required profile is empty,

unknown, or deprecated — stop.

• Acceptance ≠ exit code. A sim run / sim exec / native solver command

can return success and still be physically wrong. Always validate against an

outcome-based, bounded, measurable criterion — not "the solver ran".

• Never silently retry a failed step. Report stderr, stdout, and

run_count / completion state / relevant artifact state; let the user

decide the next move.

• Reference example values are not defaults. Values in any examples/

directory describe a specific published test case. Offer them explicitly

if useful, but wait for the user's confirmation before adopting them.

Input classification

Every task starts with: which inputs must the user supply, which may I

default, which can I derive from the files in front of me?

| Category | Rule | Examples |

|---|---|---|

| A — physical decisions | Must ask if absent. Non-negotiable. | Geometry, materials, boundary/initial conditions, physics-model choices, the acceptance criterion |

| B — operational | May default — must disclose. Affects runtime/convenience, not what the simulation represents. | --processors, --ui-mode, --workspace, smoke-test iteration counts, log verbosity |

| C — file-derivable | Infer from the actual files via a diagnostic sim exec, solver-native inspection command, or artifact parser — not from a similar example. Confirm if a downstream decision depends on it. | Mesh cell count, boundary names/types, fields present in a result file, material IDs |

Do not start until every Category A field has an explicit value from the user.

Where sim serve runs (Windows session-context foot-gun)

If you reach a remote sim host via sim --host <host> or SIM_HOST, **how

the operator started sim serve changes which drivers actually work.** This

is purely a Windows concern — Linux and macOS don't isolate display sessions

the same way.

| sim serve started from… | Headless / CLI drivers | GUI-capable drivers |

|---|---|---|

| Logged-in Windows desktop (Windows Terminal / RDP / Task Scheduler "run only when user is logged on" + interactive) | ✅ works | ✅ works — windows are visible; gui can find / click / screenshot them |

| SSH session (ssh <host> then sim serve …) | ✅ works | ❌ silent breakage — windows launch in a non-interactive session with no display surface; gui finds zero windows, screenshots come back black |

If the host advertises tools: ["gui"] but gui.find(...) returns nothing

for windows you have strong reason to believe exist, do not retry —

surface "the server may have been started from a non-interactive session" and

ask the operator to restart sim serve from a desktop session. The agent

never starts sim serve itself.

See gui/SKILL.md for the full GUI actuation API.

Long-running and multi-stage workflows

Do not wait for or request a dedicated per-task primitive (a resumable-sweep

API, an auto-coupling feature, etc.) when the generic primitives already

suffice. Bounded exec/run calls, session.health / solve-progress

inspection, and the solver's own retained project/session state are enough to

drive a long sweep, a multi-stage build, or a staged multi-physics coupling

yourself, one bounded step at a time. Reach for a new dedicated primitive only

when a real session hits a generic gap (no timeout guard, no liveness check,

no process cleanup) — not to shortcut a specific workflow that a generic loop

already covers.