CLI reference · command

bernstein run

bernstein run [OPTIONS] [PLAN_FILE]

Parse seed, init workspace, start server, launch agents. bernstein run plan.yaml # loadable YAML plan (stages + steps) bernstein conduct # reads bernstein.yaml bernstein conduct --goal "Build X" # inline goal bernstein conduct --seed custom.yaml # custom seed file bernstein conduct --plan-only # show plan without executing bernstein conduct --from-plan plan.md # execute a saved plan bernstein conduct --auto-approve # skip confirmation prompt bernstein conduct --cells 3 # 3 parallel cells (multi-cell mode) bernstein conduct --remote # bind to 0.0.0.0 for cluster access bernstein conduct --cli claude # force Claude Code agent bernstein conduct --model opus # force Opus model bernstein conduct --workflow governed # governed workflow mode bernstein conduct --routing bandit # contextual bandit routing (learns over time) bernstein conduct --routing bandit-shadow # log bandit decisions without changing live routing bernstein conduct --compliance standard # compliance mode (development/standard/regulated) bernstein conduct --container # run agents in containers bernstein conduct --sandbox docker # run agents in Docker sandbox bernstein conduct --container --two-phase-sandbox # two-phase sandboxed execution bernstein conduct --audit # SOC 2 audit mode (HMAC-chained log + Merkle seal)

Flags (28)

  • --goalTEXT

    Inline goal (skips bernstein.yaml). · flag detail

  • --seedTEXT

    Path to a custom seed YAML file (default: bernstein.yaml). · flag detail

  • --portINTEGERdefault: 8052

    Port for the task server. [default: 8052] · flag detail

  • --cellsINTEGER

    Number of parallel orchestration cells (1 = single-cell, >1 = MultiCellOrchestrator). [default: 1] · flag detail

  • --remote

    Bind server to 0.0.0.0 for remote/cluster access (default: 127.0.0.1). · flag detail

  • --cli

    [auto|claude|codex|gemini|aider|qwen] Force specific CLI agent (overrides auto- detection and config file). · flag detail

  • --modelTEXT

    Force specific model (e.g. opus, sonnet, o4-mini; overrides config file). · flag detail

  • --workflow

    [governed] Activate a governed workflow mode (deterministic phase-based execution). · flag detail

  • --routing

    [static|bandit|bandit-shadow] Model routing strategy: 'static' = fixed cascade heuristics (default), 'bandit' = contextual LinUCB bandit that learns cost- quality tradeoffs, 'bandit-shadow' = log bandit decisions without changing live routing. · flag detail

  • --compliance

    [development|standard|regulated] Compliance preset: 'development' = audit + WAL + AI labels, 'standard' = + HMAC chain + governed workflow + approval gates, 'regulated' = + signed WAL + data residency + SBOM + evidence bundle. · flag detail

  • --container

    / --no-container Run agents inside containers for kernel- level isolation (requires Docker or Podman). · flag detail

  • --container-imageTEXT

    Container image for agent execution (default: bernstein-agent:latest). Requires --container. · flag detail

  • --sandbox

    [docker|podman] Run agents in a Docker/Podman sandbox. Preferred alias for --container. · flag detail

  • --two-phase-sandbox

    / --no-two-phase-sandbox Codex-style two-phase sandboxed execution: Phase 1 installs dependencies with network access, Phase 2 runs the agent with the network fully disabled. Requires --container. · flag detail

  • --plan-only

    Generate and display the execution plan without running any agents. · flag detail

  • --from-planFILE

    Load a saved plan file and execute it (skips interactive planning). · flag detail

  • --auto-approve

    Skip the confirmation prompt before execution. · flag detail

  • --quiet

    Suppress startup/TUI output and print only the final summary. · flag detail

  • --skip-gateTEXT

    Bypass a named quality gate for this run (requires quality_gates.allow_bypass: true). · flag detail

  • --skip-gate-reasonTEXT

    Operator-visible reason recorded for quality-gate bypasses. · flag detail

  • --audit

    Enable SOC 2 audit mode: append-only HMAC- chained audit log for every task lifecycle event, with Merkle tree seal on shutdown. · flag detail

  • --ab-test

    A/B testing mode: spawn two agents with different models for each task. · flag detail

  • --dry-run

    Show scheduling plan without executing: which agent/model/tier each task would be assigned to. · flag detail

  • --profile

    Profile orchestrator execution with cProfile. Writes .prof binary and .txt report to .sdd/runtime/profiles/. · flag detail

  • -t, --taskPATTERN

    Run only backlog tasks matching PATTERN (e.g. 'gh-62' or 'mutant-fish'). · flag detail

  • --auto-pr

    Automatically create a GitHub PR when all tasks complete. · flag detail

  • --activity-logTEXT

    Write activity to log file (default: .sdd/logs/activity.log). · flag detail

  • --help

    Show this message and exit. · flag detail

Example

# print help
bernstein run --help

# common usage
bernstein run --goal VALUE