Command-line reference

The primary command is simjecture; conjecture-solver and acs are compatibility aliases.

The distribution is named simjecture. The Python import namespace remains conjecture_solver so existing capability and integration code does not need to change as part of the product rename.

uv run simjecture --help

Principal command families:

  • install: idempotently provision or verify a selected runtime profile;

  • machines: register, prepare, check and inspect local/SSH numerical workers (from 0.5.3rc2; SSH guide);

  • doctor: inspect core and optional capability health, with JSON output;

  • mvp: natural-language sandbox campaign with claims and capabilities;

  • status: compact read-only snapshot of a durable MVP run directory;

  • watch: follow durable MVP events until a terminal report or pause;

  • pause: request an action-boundary pause of a verified live runner;

  • resume: repeat a stored launch contract for a paused or incomplete run;

  • web: local research workspace and campaign monitor; the composer selects a native CLI or configured built-in API agent. --engine selects the legacy campaign engine, not the default workspace agent;

  • serve: optional hosted workspace with visitor admission, GitHub sign-in and an operator-managed queue (pip install 'simjecture[public]'). General execution requires a private Unix HTTP socket and a separately commissioned Linux broker; see server mode;

  • study: minimal, structured or frontier studies with recorded launch contracts;

  • steer: advisory guidance for an existing minimal study;

  • llm-benchmark: prepare, grade, run and publish recorded-diagnostic coding tasks (distinct from the simulation benchmark command);

  • tui: optional interactive dashboard (uv sync --extra tui);

  • benchmark: deterministic planted scientific benchmarks;

  • campaign: durable bounded campaign execution;

  • orchestrate: fixed-DAG multi-action research campaigns;

  • package verify: independently verify a discovery package;

  • schemas: export or check public JSON Schemas.

Use each subcommand’s --help output as the authoritative option reference. CLI defaults are tested and versioned with the source; documentation examples avoid duplicating the complete argument surface.

For live campaigns, prefer hypothesis and instruction files over long shell arguments so the exact operator input can be reviewed before launch. --instruction-file is accepted for the same reason. The primary Web client and maintenance-mode TUI both launch through structured input files and never build a shell command from hypothesis text.

web, status, watch, pause, and resume do not require the TUI extra. watch Ctrl-C stops the viewer only. pause never uses SIGSTOP. resume replays every structured option only for a contained, self-contained launch contract; unsafe external paths are refused. See Web interface and Terminal interface.

Native-agent studies

For a complete setup and automation recipe, see Run a headless study.

simjecture study --campaign DIR --hypothesis-file H --instructions-file I starts minimal mode by default. Add --mode structured or --mode frontier for another workflow. --backend selects codex-glm, codex, grok, agy or builtin; non-GLM backends require an explicit --model. Native CLI tools remain available. builtin uses a private --provider-config file and requires the workspace extra; it supports minimal mode only. Keep provider credentials out of committed files and shared records. --wall-seconds fixes the study budget and --turn-seconds controls the native session allowance/watchdog. Resume with the same campaign and instructions; mode and deadline are preserved. --continue-from instead creates a new minimal phase with a fresh explicit budget and selected context; see continuation and steering. simjecture-supervise and simjecture-research are equivalent launchers. The browser and TUI share the same native mode choices. Legacy mvp and DSH/API remain explicit legacy routes. See the guide.

Minimal studies can select --machine-registry DIR --machine ID, repeating --machine to choose a pool. Placement and resource requests are recorded by Lab.run; local execution remains the default when no pool is selected.

New minimal studies default to --director; --no-director disables strategy control for that launch. --judge-model and --judge-reasoning-effort choose the director and independent reviewer route. Existing studies preserve their saved policy on resume. See stop and replan.