BioScaffold OS is a sandboxed, cell-inspired software runtime simulator. The first foundation release models a single artificial cell with explicit genome, membrane, nucleus, ribosome, mitochondria, lysosome, checkpoint, audit, apoptosis, and mitosis boundaries.
The simulator does not perform uncontrolled replication, production deployment, permission escalation, audit deletion, network access, or live workflow integration. Mitosis creates a restricted child inside the local simulator only.
Install development dependencies:
python -m pip install -e ".[dev]"Run tests:
pytest -vValidate the BioComponent registry:
pytest tests/test_component_cards.py tests/test_card_registry.py -vBioScaffold exposes a noninteractive product workflow command that OpenClaw can call from the workspace:
python -m bioscaffold run-product .\request.json --output .\delivery-report.jsonMinimal request:
{
"organism_id": "organism_000001",
"product_name": "Authentication Module",
"requirements": [
{
"requirement_id": "password-policy",
"text": "Require password policy.",
"artifact_type": "code"
}
]
}If --output is omitted, the delivery report JSON is written to stdout. Reports include the terminal state, archive ref, generation IDs, task IDs, validation refs, assembly refs, immune events, and project microtask count.
For local-only autonomous project work:
python -m bioscaffold run-session .\autonomous-request.json --pretty
python -m bioscaffold session-status .\.bioclaw\sessions\<session_id>\session.json --pretty
python -m bioscaffold resume-session .\.bioclaw\sessions\<session_id>\session.json --prettyThe default runtime budget is eight hours. Autonomous mode allows local file edits, local verification commands, checkpoint writes, and local commits. By default, it denies:
- push
- deploy
- publish
- install
- destructive commands (for example, deletion or recursive removal)
- secret reads
- commands that target paths outside the workspace
For seeded workflows, pass a seed request to launch bounded generation cycles and write the summary to .bioclaw/seeds/<session_id>/seed-session.json:
python -m bioscaffold run-seed .\seed-request.json --prettyMinimal OddDB seed request example:
{
"session_id": "seed_odddb_000001",
"workspace_path": "C:/Users/jakeb/Projects/OddDB",
"organism_id": "organism_odddb_000001",
"product_name": "OddDB Autonomous Baseline",
"seed_goal": "Prepare OddDB for safe unattended BioClaw runs and verify baseline quality.",
"generation_limit": 3,
"verification_commands": [
"python -m odddb --version",
"python -m odddb doctor --json",
"python -m unittest discover -v"
]
}Seeded runs are bounded by:
generation_limit(stop after the configured number of microtask generations),max_runtime_seconds(session runtime limit for each inner autonomous session).
For full endurance runs, set:
{
"run_until_runtime_exhausted": true,
"turn_delay_seconds": 60,
"generation_limit": 10000,
"max_runtime_seconds": 28800
}In this mode a terminal/no-work repository does not stop immediately. BioClaw converts terminal state into verification-only generations and keeps running turn after turn until the wall-clock budget is exhausted, a safety policy blocks execution, verification fails, or the generation ceiling is reached.
For scheduler-owned runs, split the burn into resumable one-generation batches:
{
"resume_existing_seed": true,
"generation_batch_size": 1,
"run_until_runtime_exhausted": true,
"turn_delay_seconds": 0
}With these fields, repeated run-seed invocations append to .bioclaw/seeds/<session_id>/seed-session.json. Each batch returns status = "paused" until the generation limit, a policy failure, or the external scheduler's wall-clock deadline ends the run. Terminal/no-work repositories become verification-only generations, so the scheduler can keep burning without one long process.
Each seeded generation now records a whole-system biological fleet by default. The default fleet includes DNA, RNA, protein, membrane, mitochondria, bacteria, white blood cell, tissue, organ, organism, and generation-reviewer structures. Fleet count scales with generation index: generation 1 records 1 full fleet, generation 28 records 28 full fleets, and each fleet contains the full biological structure set. Every structure records one tiny biological action mapped to a mechanical project function in generations[].fleet_actions; runnable project work remains bounded and policy-checked through the existing autonomous session executor.
The default seeded lifecycle phase model is:
- generations 1-10:
structure_verification - generations 11-30:
work_planning - generations 31-50:
target_execution, executing the plans from generations 11-30 in order - generation 51:
cleanup - generation 52 and later:
terminal
OpenClaw fleet-turn orchestration reads that same default manifest through:
python -m bioscaffold fleet-manifest --prettyThe manifest command is the source of truth for scheduler integrations. OpenClaw uses it to create one command file per biological structure plus one fleet-turn orchestrator command file. Each OpenClaw cron turn runs that orchestrator once with a foreground exec; the orchestrator records every structure-level action, then runs one final command for the safe BioClaw generation. That keeps the local tokenless OpenClaw agent on one reliable tool call per turn while still producing a visible per-structure fleet ledger.
Seeded workflow inherits existing autonomous policy defaults:
- push is denied unless explicitly allowed,
- deploy/publish/install/destructive/path-escape/secret reads are denied,
- command/path constraints apply to every inner run.
Checkpoint data under .bioclaw/ is excluded from product commits so it is not shipped as product changes.
- Design spec:
docs/superpowers/specs/2026-05-22-bioscaffold-os-design.md - Foundation implementation plan:
docs/superpowers/plans/2026-05-22-bioscaffold-os-foundation.md