CLI Reference

CLI Reference

All commands run from mosaic-ts/ via pnpm dev <command> (development) or pnpm build && mosaic <command> (after build). Commands are registered in mosaic-ts/src/cli/index.ts; each lives in mosaic-ts/src/cli/commands/.

Output defaults to Chinese reports; CLI flags stay English. --lang zh|en|bilingual switches report language where supported, and --fake-llm is the recommended zero-cost path.

Core / plumbing

Command	Purpose
bridge-ping	Spawn the Python sidecar and verify tools.list / config.get.
tool-call <name> [argsJson]	Invoke a single sidecar tool.
tool-loop	Run the tool-report loop.

Daily cycle

pnpm dev daily-cycle --cohort cohort_default --fake-llm

Options: --cohort <name>, --date <YYYY-MM-DD>, --fake-llm, --llm-provider <name>, --model <name>, --base-url <url>, --out <path>. Runs all 25 agents through the LangGraph.js graph; the CIO writes portfolio_actions (persisted to the recommendations table).

Scorecard / Darwinian

pnpm dev scorecard --cohort cohort_default --since 2024-01-01
pnpm dev darwinian --cohort cohort_default

• scorecard options: --cohort <name>, --since <date> (YYYY-MM-DD), --out <path>. scorecard is a single view command (no subcommands).
• darwinian options: --cohort <name>, --date <YYYY-MM-DD>, --compute, --out <path>.

Forward-return back-fill is the scorecard.score_pending RPC (BridgeApi.scorecardScorePending), invoked programmatically / by the daily pipeline — it is not currently a standalone CLI subcommand. See Scorecard & Paper Trading.

Autoresearch (prompt self-evolution)

pnpm dev autoresearch trigger --cohort crisis_2008 --fake-llm --eval-days 5
pnpm dev autoresearch log --cohort crisis_2008

Subcommands: trigger, evaluate, log, branches, revert. trigger options include --cohort, --agent, --max <n>, --dry-run, --fake-llm, --eval-days <n>, --llm-provider/--model/--base-url.

Prompt Operations

pnpm dev prompts init-private-repo ~/private-mosaic-prompts
pnpm dev prompts audit-versions --status keep
pnpm dev prompts verify-release --version-id 123
pnpm dev prompts gc-worktrees --repo-target all --max-age-hours 24

• init-private-repo creates the sparse private prompt repo. --seed-baseline is migration-only and creates broad override shadowing.
• audit-versions prints metadata only: ids, hashes, repo id, status, metrics, and branches. It does not show prompt content.
• verify-release checks the pinned release tuple (code_commit_hash, prompt_repo_id, prompt_commit_hash, prompt_sha256), recomputes the prompt SHA at the commit, and runs the tool compatibility gate.
• Before release, also run pnpm prompt:drift -- --base-ref origin/main or the scheduled drift check in the private operator environment.
• gc-worktrees removes stale managed worktrees under data/worktrees for the project and/or private prompt repo.
• Private prompt repos must use a private remote with least-privilege access and encrypted backup or encrypted-at-rest storage.

PRISM (multi-regime training)

pnpm dev prism list
pnpm dev prism train --cohort crisis_2008 --fake-llm

Subcommands: list, train, status, compare. train options: --cohort/--all, --start/--end, --dry-run, --fake-llm, --max-concurrent <n>, --max-mutations <n>, LLM flags.

JANUS (cross-cohort meta-weights)

pnpm dev janus run
pnpm dev janus weights

Subcommands: run, weights, regime, history. Options: --date <date>, --window <n>, --days <n>.

MiroFish (reflexive simulation)

pnpm dev mirofish generate --swarm --seed 7      # generate scenarios
pnpm dev mirofish train --path-aware             # forward-train; --path-aware = drawdown-penalized scorer
pnpm dev mirofish history

Subcommands: generate, train, history.

• generate: --days <n>, --seed <n>, --print, --reflexive, --swarm, --engine <name>.
• train: --days, --seed, --agents <list>, --dry-run, --fake-llm, --reflexive, --engine <name>, --swarm, --scorer <name>, --path-aware, LLM flags.

Backtest

pnpm dev backtest --cohort cohort_default

Options: --cohort, --prompt-commit-hash <hash>, --fake-llm, LLM flags, --veto-threshold <num>, --initial-cash <amount>, --benchmark <ticker>, --force-refill, --log-every <n>, --out <path>. Plus backtest-fill for the cache-fill stage.

The --out flag writes the metrics JSON. Full ATLAS-isomorphic artifacts (summary.json / portfolio_trajectory.csv / equity_curve.png) are produced by the backtest.run_historical RPC when called with a results_dir (see Bridge RPC); not yet a backtest CLI flag.

Paper trading

pnpm dev paper account
pnpm dev paper buy ...

Subcommands: register, login, logout, account, buy, sell, positions, trades, suggest. See Scorecard & Paper Trading.

Data ingest (vendored qlib collectors)

pnpm dev data incremental --kind stock|etf [--end YYYY-MM-DD]
pnpm dev data validate --kind stock|etf [--gap-threshold 0.01]

Long-running (minutes) — run as cron, not alongside latency-sensitive RPCs. See Data Layer.

TUI

pnpm dev dashboard --cohort cohort_default [--user <name>]

See TUI.

Daily operation

The system is semi-automatic. A typical post-close cron pipeline:

cd mosaic-ts
pnpm dev daily-cycle --cohort cohort_default     # 25 agents → CIO portfolio (recommendations table)
# forward_return back-fill: call the scorecard.score_pending RPC (matures after T+5)
pnpm dev darwinian --cohort cohort_default
pnpm dev janus run
pnpm dev dashboard                               # review one screen

The scorecard.score_pending back-fill is an RPC rather than a CLI subcommand today; invoke it via the bridge (BridgeApi.scorecardScorePending(cohort, today)).