bmd-run

A restricted, deterministic, user-level client for bmd-compute. It is developed by the Burton Materials Discovery Lab, Tel Aviv University, and is meant for group members working in a terminal and for developer agents (Claude, Codex) that need to exercise bmd-compute the way an ordinary student would.

bmd-run is the restricted user-level client for bmd-compute. "Client" describes its architectural role, not its name: the product, repository, command, distribution and import package are all bmd-run / bmd_run.

Status: the read/build commands are a proof of concept against frozen bmd-compute v1.0.0. Milestone R1 added a separate, authenticated bmd-run api capability for one calculation at a time through BMD Compute's machine API v1 (bmd-compute e3fbb3b), and Milestone R2 wraps it in one command. See ARCHITECTURE.md for the design and its reasoning.

Running a calculation

The recommended interface is one structure, one command, one BMD Compute calculation. In a directory that contains a file named POSCAR:

bmd-run              # Energy only (BMD Compute's default Desired Output)
bmd-run --relax      # Relaxed structure
bmd-run --dos        # Electronic DOS
bmd-run --bands      # Electronic band structure
bmd-run --custom workflow.json

Or name the structure file, with optional overrides:

bmd-run Si.vasp
bmd-run Bi2Se3.cif --bands --cpus 48 --memory-gb 128 --walltime 48:00:00
bmd-run Si.vasp --desired-output energy        # the R2 form, still supported
bmd-run Si.vasp --custom-workflow workflow.json

Option	BMD Compute workflow
(none)	Desired Output energy_only
--relax	Desired Output relaxed_structure
--dos	Desired Output electronic_dos
--bands	Desired Output electronic_band_structure
--custom FILE	Custom workflow from FILE (same as --custom-workflow FILE)
--desired-output ID	any Desired Output identifier (for example phonons), or energy/relax/dos/bands

Only one workflow option may be given; combining any two (for example --relax --dos or --dos --desired-output energy) is an error and nothing is sent.

Which structure. If a structure file is named, it is used. Otherwise bmd-run uses the file named exactly POSCAR in the current directory, read as POSCAR. It does not look anywhere else (no parent or sub-directories, no CONTCAR, *.vasp or *.cif, no guessing), and it never modifies, renames or copies the file. If there is no ./POSCAR, it stops with an error before contacting BMD Compute. --format applies only to a named file. A structure that BMD Compute rejects fails at the plan step, before any attempt is recorded.

Every invocation is a new calculation. Running bmd-run twice in the same directory submits two calculations with two attempts; bmd-run does not detect repeats. To continue or check an earlier run, use its attempt ID with the api commands below, never bmd-run again.

bmd-run asks BMD Compute for the authoritative plan, prints a short summary, records the attempt locally, then has BMD Compute prepare it on POWER and submit it to SLURM. It submits without asking for confirmation: that is what the command is for. It prints the attempt ID, the SLURM job ID and how to check the job.

• Structure: a POSCAR or CIF file (.cif is read as CIF; --format overrides). The file's text is sent unchanged and the file is never modified.
• Workflow: one of the options in the table above. The values sent are BMD Compute's own identifiers; --custom/--custom-workflow take BMD Compute's Custom stage schema. With none, BMD Compute's Energy-only Desired Output is used, as in the browser interface. bmd-run infers no methodology.
• Overrides: --cpus, --memory-gb, --walltime HH:MM:SS and --queue are passed to BMD Compute as given; BMD Compute validates them and fills in everything else. bmd-run adds no methodology, defaults or resource choices of its own.
• Connection: --api-url, --token-file, --state-dir, --timeout and --json work as for the api commands below (SSH tunnel at http://127.0.0.1:18000, protected token file).

If it stops part-way. Once the attempt is recorded, bmd-run always reports its ID, how far it got and the recovery commands. Continue with the same attempt using the api commands, which re-send exactly the recorded request (a changed structure file or different options are never used):

bmd-run api prepare --attempt UUID   # if it stopped before preparation was confirmed
bmd-run api submit UUID              # submits that same attempt (never twice)
bmd-run api status UUID              # what BMD Compute reports now

A submission whose outcome is unknown (a lost response, a timeout, an unreadable answer) is reported as uncertain (exit 15), never as failed, and is never retried automatically: check it with api status or repeat api submit with the same ID. Running bmd-run again starts a new calculation with a new attempt, so do not use it to recover. If the process is killed before it can print the ID, the attempt is still recorded as <state dir>/attempts/<attempt id>.json (default ~/.local/state/bmd-run/attempts/); with --json the ID is also written to stderr as soon as it is recorded.

The api commands remain available for advanced, step-by-step use (plan only, prepare without submitting) and for recovery. For several structures, see bmd-run batch below. A structure file named like a command (plan, api, submit, ...) or like an option (--dos) must be given as ./NAME. bmd-run --help and bmd-run --version never start a calculation; bmd-run POSCAR --help (or bmd-run --relax --help) shows the options of the one-command interface.

Running a campaign (bmd-run batch, Milestone R3A)

A campaign is a small, ordered list of single calculations in a JSON manifest. Each one is run exactly as bmd-run STRUCTURE would run it (same request builder, same plan, record, prepare and submit operations), and each is bound to exactly one attempt that is reused whenever the batch is run again.

bmd-run batch benchmark.json               # submit every calculation in the manifest
bmd-run batch benchmark.json --max-new 2   # testing: start at most 2 new attempts this run
bmd-run batch status benchmark.json        # read-only report, refreshed from BMD Compute
bmd-run batch status benchmark.json --local   # report from the local records only

Batch manifest

{
  "version": 1,
  "campaign": "dft-benchmark-1",
  "calculations": [
    {"cell": "si-energy-24", "structure": "structures/Si.POSCAR", "desired_output": "energy_only"},
    {"cell": "si-energy-48", "structure": "structures/Si.POSCAR", "desired_output": "energy_only",
     "cpus": 48, "memory_gb": 64, "walltime": "12:00:00"},
    {"cell": "nio-dos", "structure": "structures/NiO.POSCAR", "desired_output": "dos"},
    {"cell": "si-custom", "structure": "structures/Si.cif", "custom_workflow": "relax_static.json"}
  ]
}

• version must be 1; campaign and every cell are identifiers of 1-64 letters, digits, ., _ or - (starting with a letter or digit). Cell identifiers are unique within the manifest, ignoring case.
• structure (required) is a POSCAR or CIF file; format (poscar or cif) is optional as for --format. custom_workflow is a Custom workflow JSON file. Relative paths are resolved from the manifest's directory, not the current one.
• Exactly one of desired_output (an identifier or energy/relax/dos/bands) or custom_workflow. There is no default in a manifest.
• Optional cpus and memory_gb (positive JSON integers), walltime (HH:MM:SS) and queue, passed to BMD Compute unchanged. No other keys are accepted; BMD Compute resolves everything else.
• Any number of calculations (the manifest file may be up to 64 MiB). The whole manifest is read and every request is built before anything is sent; any problem stops the batch with nothing sent. Calculations are submitted one after another; SLURM schedules them.

What a run does

For each calculation, in manifest order:

1. A cell that has never run is recorded locally first (<state dir>/campaigns/<campaign>/<cell>.json, private), with the attempt UUID it will use, before that UUID is sent anywhere. Every remaining cell is started, unless --max-new N (for testing) limits this run to N new attempts.
2. The attempt is then advanced with the R1 operations: plan, record the attempt, Prepare, Submit. The cell keeps the plan digest and the plan's structure summary, resolved workflow and resources.
3. A cell whose attempt is already submitted is left alone (nothing is sent). A cell whose submission is uncertain keeps its attempt and is never submitted again by the batch; check it with bmd-run api status UUID (shown in the output).

Restarting is safe. Running the same command again continues the campaign: recorded cells reuse their attempt (an unprepared one is prepared, a prepared one is submitted), and the remaining cells start. A recorded cell whose structure file, workflow or resources have changed is refused, with nothing sent; give a changed calculation a new cell identifier. New calculations may be appended to the manifest at any time.

One run per campaign. While a bmd-run batch is advancing a campaign it holds an exclusive lock on that campaign's state (<state dir>/campaigns/<campaign>.lock, a POSIX flock). A second bmd-run batch for the same campaign, through any manifest file, stops at once with exit code 21 and sends or records nothing. The lock is released when the run ends, however it ends (including Ctrl-C or the process being killed). Other campaigns run independently, and batch status takes no lock, so it can be used while a campaign is running.

When it stops. A problem with one calculation (BMD Compute rejects its structure or workflow, its plan changed, preparation or submission definitely failed, or its submission outcome is uncertain) is recorded on that cell and the batch moves on to the independent cells. Anything campaign-wide stops the batch at that cell and leaves the rest untouched: a lost connection or timeout, a credential or scope problem, a local state problem, Ctrl-C, or any other error not specific to one cell. The exit code is that of the first error (see Exit codes); run the batch again to continue. The client never retries on its own.

Status

bmd-run batch status MANIFEST sends only GET /api/v1/attempts/{uuid} for recorded attempts (it never plans, prepares or submits) and classifies each cell:

Status	Meaning
pending	Not started, or not yet submitted (it will continue on the next run)
submitted	Submitted; the scheduler reports it queued or running
completed	The scheduler reports it finished successfully
failed	The job ended unsuccessfully, or BMD Compute rejected the calculation (needs a change)
uncertain	The submission outcome is unknown; check it with the commands shown

The table and --json output carry what a benchmark needs per cell: campaign and cell identifiers, attempt UUID, plan digest, structure (formula, atoms, space group), resolved workflow (Desired Output and stages), nodes, CPUs, requested memory, walltime limit, partition, SLURM job ID, and the scheduler's state, elapsed time and exit code as reported by BMD Compute. Memory is the memory requested for the job (requested_memory_gb); it is not measured peak memory, and bmd-run does not measure memory use.

--json writes one schema: "bmd_run.batch_output", schema_version: 1 document (src/bmd_run/schemas/batch-output-v1.schema.json) for both commands, with the same rules as the other outputs: no Compute text, never the token.

What it can do

Command	What it does	Compute requests
STRUCTURE	Plan, record, prepare and submit one calculation (see above)	POST /api/v1/plans, PUT /api/v1/attempts/{uuid} (twice)
batch MANIFEST	The same, for each calculation of a campaign, reusing recorded attempts (see above)	the same, per cell
batch status MANIFEST	Read-only campaign report	GET /api/v1/attempts/{uuid} per recorded cell
identity	Behavioural fingerprint of the service. Not verified source identity.	GET /openapi.json, POST /analyze, POST /build-workflow (fixed probe)
options	Desired Outputs, stage types, theories and modifiers Compute offers	POST /analyze (fixed probe)
analyze FILE	Compute's structure summary, method considerations and default workflow	POST /analyze
plan OUTPUT FILE	Compute's resolved workflow and resources, then consideration status, structure facts, and per-stage INCAR/KPOINTS rebuilt by the client	POST /build-workflow

The separate api commands (Milestone R1, see below) use BMD Compute's authenticated machine API:

Command	What it does	Compute requests	Token scope
api plan FILE	The authoritative plan and its digest. No side effects.	POST /api/v1/plans	plan
api prepare FILE	Plan, record a new attempt locally, prepare it on POWER (submit=false)	POST /api/v1/plans, PUT /api/v1/attempts/{uuid}	plan, prepare
api prepare --attempt UUID	Repeat Prepare for a recorded attempt, from its record only	PUT /api/v1/attempts/{uuid}	prepare
api submit UUID	Submit that same attempt (submit=true, nothing else changed)	PUT /api/v1/attempts/{uuid}	submit
api status UUID	The attempt's state, job ID and scheduler state	GET /api/v1/attempts/{uuid}	read

What it cannot do

It has no implementation of, and no hidden path to:

• SSH, SFTP or SLURM (BMD Compute alone connects to POWER and submits);
• creating SSH tunnels, running processes or remote commands;
• Monitor, Resume, cancellation or result retrieval;
• Materials Project retrieval, resource prediction or adaptive experimental design (a batch runs exactly the manifest it is given);
• Git or MongoDB;
• arbitrary HTTP.

The read/build commands cannot prepare or submit. The complete list of requests the client can make is two closed tables. The frozen v1 table (src/bmd_run/endpoints.py, unchanged):

GET  /openapi.json
POST /analyze
POST /build-workflow

and the authenticated machine-API table (src/bmd_run/api_endpoints.py):

POST /api/v1/plans
PUT  /api/v1/attempts/{canonical lowercase UUID}
GET  /api/v1/attempts/{canonical lowercase UUID}

It contains no scientific methodology and does no scientific validation. bmd-compute decides what is valid. The client only checks the shape of its own arguments: identifiers, file size, UTF-8 text, and resource values as short tokens.

Install and run

Standard library only, Python 3.10 or newer:

pip install .            # or: PYTHONPATH=src python -m bmd_run ...
export BMD_COMPUTE_URL=http://<compute-host>:<port>

bmd-run identity
bmd-run options
bmd-run analyze POSCAR
bmd-run plan energy POSCAR
bmd-run plan dos POSCAR --json
bmd-run plan relax structure.cif --cpus 48 --walltime 24:00:00

The plan aliases are spelling conveniences for Compute's own Desired Output identifiers:

Alias	Desired Output
energy	energy_only
relax	relaxed_structure
dos	electronic_dos
bands	electronic_band_structure

Any other identifier is sent as typed, and Compute accepts or rejects it. Under v1.0.0, dos resolves to PBE geometry optimisation → HSE06 static → HSE06 DOS, which is why plan prints the resolved workflow and resources first.

Authenticated machine API (bmd-run api, Milestone R1)

bmd-run on POWER -> HTTP over your existing SSH tunnel -> bmd-compute on the VM -> Compute-owned POWER execution

bmd-run stays a client. BMD Compute alone decides methodology, workflow construction, automatic treatments, resources, remote preparation and SLURM submission. bmd-run sends only what you give it: the structure file's text (read, never modified), a Desired Output identifier or a Custom workflow in Compute's stage schema, and optional --cpus, --memory-gb, --walltime, --queue. It fills in no defaults and infers nothing.

# once: store the token BMD Compute issued you, readable only by you
mkdir -p ~/.config/bmd-run && install -m 600 /dev/null ~/.config/bmd-run/api-token
cat > ~/.config/bmd-run/api-token        # paste the token, then Ctrl-D

bmd-run api plan POSCAR --desired-output energy_only          # review the plan and its digest
bmd-run api prepare POSCAR --desired-output energy_only \
        --expect-plan-digest sha256:...                         # optional: bind to the reviewed plan
bmd-run api submit 6f1d2c3b-...                                 # the attempt ID prepare printed
bmd-run api status 6f1d2c3b-...
bmd-run api plan structure.cif --custom-workflow stages.json  # {"stages": [{"stage_type": ..., "theory": ...}]}

Transport and credentials. The default API origin is http://127.0.0.1:18000, the POWER end of the existing SSH tunnel (--api-url or BMD_RUN_API_URL to change it). Credentials are sent only to http://127.0.0.1:PORT or http://[::1]:PORT; host names such as localhost and every other plain-HTTP host are refused, and HTTPS is refused unless BMD_RUN_API_ALLOW_REMOTE_HTTPS=1 records a reviewed deployment (certificates are always verified). The token is read from --token-file, else BMD_RUN_API_TOKEN_FILE, else ~/.config/bmd-run/api-token (a regular file owned by you, mode 600), or from BMD_RUN_API_TOKEN. It goes only into the Authorization header: never into URLs, output, logs or attempt records. bmd-run never opens SSH connections, tunnels or processes itself.

Attempts and recovery. api prepare asks Compute for the plan, records its digest, chooses a UUID, and writes the attempt record (mode 600, under --state-dir, BMD_RUN_STATE_DIR, $XDG_STATE_HOME/bmd-run or ~/.local/state/bmd-run, directories mode 700) before sending Prepare. The record holds the attempt UUID, the Compute API origin, the exact request and its SHA-256, the expected plan digest and the locally observed state; never credentials, SSH settings, remote paths or submission identity tokens.

• api submit and api prepare --attempt re-send the recorded request; they take no structure, workflow or resource arguments, so changed CLI defaults can never change a resumed request. A record whose request no longer matches its stored SHA-256 (accidental corruption, a partial edit) is refused. This is a consistency check, not tamper protection: a record deliberately edited and rehashed by its owner is not guaranteed to be detected, and deliberate local modification is outside the threat model. Once Compute has registered an attempt UUID, Compute's binding of that UUID to its request is authoritative and a changed request is refused (exit 14).
• After a timeout or lost response, repeat the same command with the same attempt ID, or run api status. A new UUID is never generated automatically.
• An uncertain submission (exit 15) is never answered by another attempt. BMD Compute never submits one attempt twice.
• api status UUID also reads attempts created by other clients under the same Compute principal, without a local record (nothing is written for them). Compute's optional campaign/cell labels are validated against Compute's label rule; only the label names are shown, never their values.

--json output (for synthetic-user testing)

bmd-run [STRUCTURE] --json writes one schema: "bmd_run.run_output", schema_version: 1 document (src/bmd_run/schemas/run-output-v1.schema.json), on success and on error. It reports stage (plan, prepare, submit or complete), attempt_id once the attempt is recorded, the recovery commands, and in result the plan and attempt objects of bmd_run.machine_output v1, under the same rules.

Every command accepts --json, before or after the command name. It then writes exactly one JSON document to stdout, on success and on error, conforming to src/bmd_run/schemas/output-v3.schema.json (schema: "bmd_run.output", schema_version: 3).

Consumers should:

• check schema and schema_version first;
• then check ok;
• then read result (on success) or error (on failure).

compute.adapter_status is "transitional" while the client reads Compute v1's HTML pages.

Every object in the schema is closed (additionalProperties: false), and vocabulary fields are enums. Schema versions 1 and 2 relayed Compute text. They were never released and are superseded.

The api commands write schema: "bmd_run.machine_output", schema_version: 1 (src/bmd_run/schemas/machine-output-v1.schema.json) with the same rules: no Compute prose (error messages, suggestions, option text, module names, the space-group symbol) is relayed; vocabulary values are the client's copies of a closed vocabulary pinned to bmd-compute e3fbb3b (api_vocabulary.py); digests, job IDs and timestamps are re-rendered after strict pattern checks.

No Compute text reaches output

Compute's responses are untrusted, and the client relays none of their text. Pattern-based secret screening was tried and abandoned: split or re-encoded secrets passed through free-text fields. Instead, the client never relays Compute-originated text at all. Every string in a result, and every word of the human-readable output, is one of:

1. a client constant;
2. the client's own copy of a member of a closed vocabulary pinned to BMD Compute v1.0.0 (v1_vocabulary.py): Desired Outputs, stage types, theories, modifiers, VASP executables, partition/account, consideration ids and their exact v1 titles, failure titles, crystal systems, element symbols;
3. rendered by the client from numbers parsed within bounds (at most 12 significant digits), for example lattice parameters, resources and INCAR values.

What each command reports:

• Workflow: Desired Output and per-stage type, theory and modifiers. Stage names are rebuilt by the client from those values.
• Resources: typed values.
• Structure facts: formula rebuilt from element symbols and counts, atom count, volume, density, lattice, space-group number, crystal system.
• Method considerations: id and a status (applied / not_applied_automatically / advisory / unrecognized), recognised from v1's exact titles. The prose is not read.
• Generated inputs: INCAR and KPOINTS for every stage, rebuilt from validated tokens (v1_inputs.py).
  • INCAR tags come from a closed tag vocabulary generated from pymatgen's incar_parameters.json (incar_tags.py, made by tools/generate_incar_tags.py). Values are parsed by syntactic type, and word values must be from that tag's closed list.
  • Unknown and free-text tags (e.g. SYSTEM) are withheld: counted, never shown.
  • KPOINTS automatic meshes are shown in full. Explicit lists are reported as a count only.
• Compute rejections: a stage category (structure_validation, calculation_validation, unrecognized) and fixed client text.

Fields v1 offers only as free text are listed in each result under unavailable_from_v1_machine_interface:

• consideration and error prose;
• display names and calculation-plan text;
• the space-group symbol;
• comments;
• explicit k-point coordinates;
• the POSCAR preview;
• scripts and paths.

They return with /api/v1.

Residual channel (accepted, documented): a deliberately malicious Compute could encode data in the numbers it emits. Each number is bounded and re-rendered by the client, so this is a low-bandwidth numeric covert channel, not a text relay. Closing it entirely would mean reporting no Compute-derived numbers.

Closed endpoint table

Endpoint members are opaque identifiers with no route attributes, and assigning to them raises an error. Routes are immutable NamedTuples in a read-only mapping read only by endpoints.resolve(). The transport then re-checks the resolved method and path against a literal allowlist before sending anything.

Exit codes

Code	Meaning
0	Success
1	Compute rejected the request (category in error.compute_stage; Compute's prose is not relayed)
2	Usage error; nothing was sent
3	Compute unreachable
4	Unexpected response / contract mismatch
5	identity --strict fingerprint mismatch
10	api: authentication failed (token not accepted)
11	api: the token lacks the scope this command needs
12	api: invalid structure, workflow, resources or request fields
13	api: plan digest mismatch (the plan changed; nothing was prepared or submitted)
14	api: attempt conflict (bound to another request, principal or origin)
15	api: submission uncertain; check with api status, never create a new attempt
16	api: quota exceeded
17	api: service unavailable (tunnel down, Compute busy or not configured)
18	api: network timeout
19	api: no usable API token
20	api: attempt not found on Compute
21	api: local attempt state missing, unsafe, damaged or inconsistent
22	api: remote preparation or submission failed on POWER

Tests

python -m unittest discover -s tests

The suite includes:

• the static security checks;
• endpoint-mutation regressions;
• the regressions for both Codex findings: the whole token in workflow_spec_json, and the split-token and split-UUID attacks;
• an adversarial matrix that injects the page's real token and attempt UUID, in 11 encodings, into every Compute field the adapter still reads and into fields it ignores. The encodings are plain, split with spaces or newlines, UUID groups, HTML entities, percent-encoded, reversed, hex, base32, interleaved with client words, and zero-width-joined.
• an installed-wheel smoke test that builds from a clean source copy, installs into a fresh environment, checks bmd-run --help / --version, imports bmd_run, and inspects wheel contents;
• for bmd-run api: a fake BMD Compute served over real HTTP on 127.0.0.1 from responses recorded from bmd-compute e3fbb3b with Compute's own in-memory fake POWER (tests/fixtures/compute_api_v1/, tools/record_machine_api_fixtures.py). It covers POSCAR, CIF, Desired Output and Custom requests, scopes, the endpoint allowlist, persistence before Prepare, Prepare -> Submit, repeated and interrupted requests, changed-request rejection, malformed responses and token redaction, plus an information-flow matrix over every recorded response field.

The matrix checks an information-flow property rather than looking for secrets. Every alphabetic word in JSON and human output must be a word owned by the client package (numbers and client-computed SHA-256 digests excepted), and no payload fragment may appear.

Optional live run against a clean local Compute v1.0.0 (Analyze/Build only, no remote side effects):

BMD_RUN_LIVE_URL=http://127.0.0.1:8765 \
  python -m unittest discover -s tests -p "test_live*.py"

Fixture provenance

tests/fixtures/compute_v1.0.0/ holds verbatim responses from a clean local Compute v1.0.0. tools/fixtures_v1.py re-captures them through the client's own closed transport, with Compute used only as an external HTTP service, and compares them with the committed files. Only four documented per-request values are normalised (identity token, UUIDs, run timestamps, exact-script SHA-256). See tests/fixtures/compute_v1.0.0/RECORDING.md.

Transitional adapter

Compute v1.0.0 has no machine-facing API, so v1_html.py reads the browser template from commit a746155. It is intentionally minimal and will be replaced by Compute's planned JSON /api/v1 facade.

License and security

bmd-run is released under the MIT License. See LICENSE.

Security-sensitive reports should be made privately first. See SECURITY.md. Small contributions that preserve the client boundary are welcome; see CONTRIBUTING.md.