openai/boltz-structure-and-binding

Workflow

If boltz-api is missing from PATH, use boltz-cli-setup for install/update guidance before retrying. If a command reports missing or expired authentication, use boltz-cli-setup to start boltz-api auth login --device-code before retrying; do not ask permission first. If the agent host sandbox blocks boltz-api install/auth/API calls, use boltz-cli-setup to request the host sandbox bypass/escalation needed for user-wide CLI install, browser login, credential storage, temp files, or API access before retrying.

Use this skill for one defined complex, not a library workflow.

1. Normalize the inputs into entities. Each entity is {type, chain_ids, value} — note plural chain_ids (an array, even for one chain) and the field is value, not sequence:

   {"entities": [{"type": "protein", "chain_ids": ["A"], "value": "MKTAYIAKQRQISFVKSHFSRQ"}]}
   

   type is one of protein | rna | dna | ligand_smiles | ligand_ccd. Chain IDs go in entity order (A, B, C, …) unless the user specifies otherwise. Read references/api.md for per-type field variants (cyclic, modifications, ligand CCD codes, etc.) before authoring your first payload — agent guesses like sequence: or chain_id: "A" (singular) fail with unclear 400 errors.

2. If the user wants binding metrics, add a flat binding block with an explicit type field. For ligand-protein binding use:

   binding:
     type: ligand_protein_binding
     binder_chain_id: B
   

   For protein-protein binding use:

   binding:
     type: protein_protein_binding
     binder_chain_ids: [B]
   

   Do not nest the variant name under binding (for example, no binding.ligand_protein_binding object).

3. Supported optional features include constraints, bonds, modifications, model_options, and binding metrics; only add them if the user asks. Read references/api.md for exact shapes and examples.

4. Author the payload YAML or JSON, run estimate-cost, show the USD cost, wait for explicit confirmation.

5. start to submit (synchronous). Capture the ID.

6. Launch download-results with the agent runtime's background/non-blocking command facility so polling + download continue without blocking the agent session. In Claude Code, use Bash with run_in_background: true. In Codex, run download-results as a foreground shell command with yield_time_ms: 1000; if Codex returns a session_id, keep it for optional same-thread polling, but treat download-status plus the run directory as the durable source of truth. In Codex app/desktop runtimes that expose same-thread heartbeat automations, create a heartbeat that checks download-status periodically and posts a concise completion or failure update when the download reaches a terminal state. After launching the downloader, always report the job ID, run name, and output directory. Include the next check cadence if the heartbeat was created; otherwise include the download-status command.

Command Pattern

# Replace placeholders with concrete absolute paths before running.
# Use a short descriptive run name, for example: sab-<target>-<ligand>-v1

# 1. estimate
boltz-api predictions:structure-and-binding estimate-cost \
  --model boltz-2.1 \
  --input @yaml:///absolute/path/payload.yaml

# 2. confirm with user, then submit
boltz-api predictions:structure-and-binding start \
       --model boltz-2.1 \
       --idempotency-key "<run-name>" \
       --input @yaml:///absolute/path/payload.yaml \
       --raw-output --transform id

# 3. Copy the printed job ID into this command, then launch it in the agent
# runtime's background/non-blocking mode.
# Claude Code: Bash with run_in_background=true.
# Codex: foreground shell command with yield_time_ms=1000; keep the returned session_id if one is provided.
# Do not append "&" or use nohup in Codex.
boltz-api download-results \
  --id "<job-id-from-start>" --name "<run-name>" \
  --root-dir "/absolute/path/boltz-experiments" \
  --poll-interval-seconds 10
# -> /absolute/path/boltz-experiments/<run-name>/outputs/archive.tar.gz, .boltz-run.json

Always Do This

• Keep payload field names exactly as the API body names shown in references/api.md; then pass the merged payload with --input @yaml:///absolute/path/payload.yaml or @json:///absolute/path/payload.json. Never use @./payload.yaml or @file:// for object-typed payloads.
• Use absolute paths for the output root, payload files, and embedded structure files. Do not cd into the run directory for follow-up commands; pass the same --root-dir and use absolute paths so later relative paths do not drift.
• Residue indices are 0-based wherever the payload asks for residue positions (constraints, modifications, contact tokens).
• For CIF/PDB bytes embedded in --target / structure.data, use @data:///absolute/path/file.cif — it detects binary and base64-encodes. Don't use bare @path for binary data.
• Use the same slug as both --idempotency-key at submit time and --name at download time so re-runs are idempotent and resume from .boltz-run.json.
• In permission-gated agents such as Claude Code, keep each Boltz call as a top-level command that starts with boltz-api. Prefer concrete arguments over sh -c, inline environment assignments, aliases, wrapper scripts, loops, or pipelines around the boltz-api invocation unless the user already allowed that exact command form. Use --raw-output --transform id, read the printed ID, then paste that literal ID into the next download-results command.
• Prefer the agent runtime's background/non-blocking command mode for download-results. In Codex specifically, keep download-results in the foreground and set the shell tool yield to 1000 ms; Codex will return a session_id if the command is still running. Do not append & or use nohup in Codex because the tool runner may clean up shell-backgrounded descendants before .boltz-run.json is fully written.
• After the background/session starts, do not manually wait on it or run ad hoc polling loops. download-results emits JSONL progress on stderr by default; add --progress-format text --verbose only when you explicitly want human-readable logs.
• In Codex app/desktop runtimes with same-thread heartbeat automation support, schedule a heartbeat after launching download-results. The heartbeat should run boltz-api --format json download-status --name "<run-name>" --root-dir "/absolute/path/boltz-experiments" on a sensible cadence, post only material status changes or terminal completion/failure, and stop once terminal. Poll the saved session_id with an empty write_stdin only for interactive, user-requested progress checks. Do not loop retrieve yourself.
• If the current host has no heartbeat automation support, do not claim an automatic next check. Report the job ID, run name, output directory, and the command needed to check download-status.
• If detached download needs to be restarted, re-run boltz-api download-results with the same --name "<run-name>" and the same --root-dir.
• Poll interval: keep --poll-interval-seconds 10 for SAB — predictions usually finish in under a few minutes.
• Cost: there is no published per-unit rate to cite for SAB — run estimate-cost and state only the figure it returns. Don't estimate or comment on cost.

Escape Hatch

For anything not covered in references/api.md:

• Payload reference: https://api.boltz.bio/docs/api/python/resources/predictions/subresources/structure_and_binding/methods/start
• CLI flag names: boltz-api predictions:structure-and-binding start --help (schema details aren't there — just flag names and types)

Read references/api.md for entity shapes, binding variants, bonds, constraints, model options, and input examples. Read references/results.md when summarizing downloaded outputs, metrics, or validation quirks.

Outputs

Summarize metrics.json and point the user at the downloaded CIF path. Read references/results.md for the local layout, nested metrics, binding metric variants, and SAB validation quirks.

SAB 400 validation quirk

If the server rejects a payload with only {"code":"VALIDATION_ERROR","message":"Request validation failed"}, inspect entities, binding, and constraints; read references/results.md for details.