oro-env-runtime

Portable sealed-pack execution and verification for ORO environments.

The package owns the public runtime contract used by environment generators and validators:

• sealed-pack schemas, versions, integrity checks, and portable validation;
• pack-pinned, index-backed catalog and local-search access;
• TaskSession, environment execution, replay, and user simulation;
• deterministic verification, rewards, and runtime family behavior.

It deliberately excludes private catalogs and task packs, seed search, task-family generation, catalog profiling, compiler orchestration, model configuration, and deterministic recompilation.

Requires Python 3.11 or newer and is distributed under the MIT License.

Pack boundary and compatibility

The runtime consumes a sealed directory with exactly manifest.json, checksums.sha256, tf4_hybrid_release_gate.json, and data/tasks/private_tasks.jsonl. Catalog source files and the Lucene index stay outside the archive. The manifest pins their hashes and the search-server identity that must be available when the runtime validates or opens the pack.

This source revision supports pack format oro_compiled_epoch_v5, environment contract oro.env.v3, runtime contract 0.3.2, tool contract oro_task_tools_v5, verifier 0.3.4, and result schema v2. It also requires the exact event and replay contract versions exported from oro_env_runtime.contracts. These values are execution compatibility identities. They are not the same namespace as the Python distribution version in this package's pyproject.toml.

validate_epoch() fails closed unless the pack file set and checksums are exact, every manifest contract equals the runtime constants, the search server matches the sealed index identity, and the task schemas, catalog references, admission evidence, counts, fingerprints, and TF4 release gate are valid. TaskSession validates an uncached pack and stores a process-local seal for that path. For performance, each file digest is reused while its size and mtime_ns are unchanged. This cache invalidation is not an adversarial tamper-detection boundary because a same-length mutation can restore the original timestamp. Consumers must extract or mount each pack on a read-only or otherwise immutable filesystem and prevent concurrent mutation for the full execution lifetime. Within that trust boundary, TaskSession compares the seal before and after construction and revalidates when an ordinary file change alters the observed metadata.

The control plane may derive a qualifying or race delivery subset with build_delivery_subset_archive(). The helper preserves the parent pack as immutable provenance, selects an exact task roster, updates the manifest aggregates and task-set fingerprint, regenerates the checksums, and emits deterministic archive bytes with their own SHA-256. Delivery manifests carry an explicit scope, parent hash, and roster. validate_epoch() applies the same task, catalog, verifier, and checksum checks as a full pack while allowing a race subset to have non-uniform family counts; validate_delivery_binding() additionally binds those valid bytes to the scope and roster authorized by the control plane. A parent artifact signature does not sign derivative bytes and must not be presented as their signature.

Policy boundary and state visibility

TaskSession.policy_view() returns the task query, local tool schemas, and contract versions. TaskSession.step() and step_parallel() return only observation, done, and error. The complete immutable ledger stays inside the session for replay and verification.

State has these visibility classes:

• Public: the goal, budget, frozen catalog facts, cart, tool schemas, policy errors, and terminal status.
• Private: gold_set, acceptance, admission evidence, verifier checks and rewards, the complete ledger, verifier signals, and true-state hashes.
• Conditionally revealed: the market-event kind and delta. The environment applies the event to the first product that the cart accepts. The solver already supplied this product. The environment reveals the event only in the response to that successful add_to_cart call.

Before and after the trigger, a change to a private field must not change any policy-visible value when public state, seed, and actions are equal. The event target comes from the public action, not from acceptance, gold, admission, or verifier state. Acceptance, gold, admission, verifier, ledger, and true-state hash data stay private. A task without an event has no conditionally revealed state. The sealed-pack validator records its event noninterference check as not_applicable.

For an event task, the first successful add is also the family-owned public commitment used by the reward contract. Exploratory adds remain valid environment actions, but an add before the intended commitment makes the event construct unsuccessful.

Shopper interaction protocol (oro_task_tools_v5)

TaskSession.run() owns the managed solver and shopper-simulator loop. The message tool may be called once or alongside other tools in an ordered solver turn. After the group finishes, the loop may append one simulator reply to the solver transcript and records it as an authoritative user_message ledger entry. A group containing multiple message calls still produces one reply to the complete turn.

The managed loop may add one unsolicited shopper reply after a visible market event. It deliberately waits for one unaided solver turn after the event fires, then appends the reply to the transcript and ledger if the session remains active and the task is not state-blind. It does not push for arbitrary policy actions. A terminal action during the unaided turn ends the episode without simulator rescue.

The lower-level TaskSession.step() and step_parallel() APIs do not run the shopper simulator. They execute one ordered action group and return only observation, done, and error for each call. Later calls in a group return skipped_after_terminal after a terminal action and cannot change state. policy_view() advertises oro_task_tools_v5 and result schema v2; verdict() returns the typed VerifierResult. The immutable ledger and verifier internals are not included in policy-visible step results.

The simulator has dialogue authority only. It cannot mutate commerce state or assign reward. Provider timeout, unavailability, or invalid output is an environment failure and quarantines the session; it is never classified as miner failure. Episode evidence records each delivered message and the simulator model, prompt, configuration, normalized response, latency, and sanitized failure type. Credentials, provider exception detail, hidden task fields, verifier state, and reward remain private.

Scripted intervention delivery

A multi-call solver group is one atomic turn. The harness delivers all due scripted interventions after a nonterminal group, before the next solver turn. A group containing add_to_cart followed by place_test_order can finish before a newly due cart-count intervention is delivered. The harness does not deliver updates after termination or apply an undisclosed intervention budget to an earlier order.

Adapters use oro_env_runtime.loop.due_interventions to select due updates and oro_env_runtime.loop.record_user_message to record delivered shopper decisions in the authoritative ledger. The private _record_user_message name remains available for validators using the 0.2.7 interface.

Releasing

Bump project.version in this package's pyproject.toml as part of a reviewed pull request. When that change reaches main, GitHub Actions creates the matching runtime-vMAJOR.MINOR.PATCH tag and dispatches the protected publish workflow at that tag. The publish workflow reruns the generator tests and lint, builds and audits the public wheel, verifies it in a clean Python 3.11 environment, and publishes it to PyPI through Trusted Publishing.

PyPI releases are immutable. Fix a bad release with a new version rather than reusing a tag or overwriting an existing artifact.