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;

  • deployment: plan or check a versioned campaign deployment without installing software or starting a study (deployment manifest);

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

  • watch: follow durable research 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;

  • 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 research with recorded launch contracts;

  • attention: inspect optional Conjecture 1 scientific-attention tooling, write an example config, verify a resident model service or score an observation (scientific attention);

  • 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);

  • 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.

simjecture study limits --campaign PATH inspects operating allowances. Add --limits-file FILE --reason TEXT to update them without changing the scientific contract or campaign deadline. The same control is Run settings in the browser; see operating settings.

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

web, status, watch, pause, and resume are included in the core package. watch Ctrl-C stops the viewer only. pause never uses SIGSTOP. resume preserves the saved minimal launch contract and absolute deadline. Classical, DSH, structured and frontier runs cannot be resumed in this release. See Web 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 a minimal study. Minimal is the only supported 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. The built-in API agent uses smolagents directly. API output-token budgets use provider defaults unless the private provider profile explicitly sets max_tokens. API research workers retain the preceding turn’s structured tool calls and results alongside the current evidence brief. Their action limit creates an incomplete checkpoint, not a research conclusion or an extra model-generated summary. Older evidence remains in project files, receipts and the research journal; independent reviewers receive separate contexts. --wall-seconds fixes the study budget and --turn-seconds controls the native session allowance/watchdog. Set --max-experiment-bytes and --max-total-bytes for the initial per-experiment and aggregate logical storage allowances; the total must cover at least one experiment. Set --report-reserve-seconds to a positive value smaller than the wall budget to reserve time for final reporting. Omitted options retain existing defaults. Resume with the same campaign and instructions; initial allowances, report reserve, mode and deadline are preserved. Explicitly changing an initial allowance or report reserve on resume is rejected. Storage allowances are logical limits, not a guaranteed physical free-space reserve. --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 also launches minimal studies. Removed workflows are described in the migration guide.

For an active minimal study, simjecture study budget --campaign DIR reads the authoritative clock and simjecture study extend --campaign DIR --hours 12 --request-key KEY grants additional time without restarting the supervisor. See live budget extensions for the operator API, agent requests, audit records and individual experiment limits.

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.