Start an agent on one bounded task

This is the initial agent entry point for the MCRP research seed. It accepts a local JSON task and returns machine-readable accepted records or refusals. No human onboarding session, account or network service is needed. Python3.9+ and the standard library are sufficient.

From the repository root:

python3 papers/07-release-packet/agent_api/cli.py \
  --policy papers/07-release-packet/agent_api/examples/policy.json \
  --request papers/07-release-packet/agent_api/examples/physics_astro.json

Read manifest.json for machine discovery, the request schema for structure, and the four complete JSON fixtures under examples/. Relative paths in the manifest are relative to this directory unless labeled as repository-root commands. The CLI never fetches the schema URL or any URL in an evidence record.

Each example should return completed_with_refusals: 12 submitted commands, 11 accepted, one intentional stale-check refusal, two exact offers, two checks and two scoped reliance records. The material correction supplies genuinely changed claim/evidence content and a new version. The new version’s reliance is current under the fixture while the old reliance remains reconsideration-pending. A new support assertion does not silently repair old scientific content.

The four examples cover calibration arithmetic for physics/astronomy, an assay mean for biology, a completion denominator for economics/social science, and a fictional event ordering for law. These are small calculations, not domain validation, biological efficacy, causal estimates or legal advice. The check method and result are caller assertions; this CLI does not execute a scientific analysis or establish that the assertion is true.

The small interface

A task contains schema, synthetic_fixture: true, task_id and commands. Each command has an immutable local id, an action, asserted agent_id, integer simulation now, and input. The four main actions are offer, check, rely and amend. Observe queries a receipt’s local currentness. Availability records a local evidence-access condition. Adapt_legacy exposes the existing adapter’s explicitly non-authorizing projection and loss report.

Command aliases are local to one batch, unique and type checked. Only successful offer/check/rely commands create referenceable records. A failed future-dated command cannot advance the clock, consume check capacity or create an alias. Successful commands commit one at a time; earlier successful commands are retained when a later command is refused. This is a serial in-memory transaction boundary, not concurrent storage or durable delivery.

Agent identity, principal and control are different

The separate trusted policy file maps agent IDs to principals, and principals to control groups, skills, roles, operations and declared check budgets. The example review-agent and review-team-child share one principal. Both debit the same principal’s check-effort budget, and they do not become two independent control groups. Two different principals may also share a control group and therefore not supply independent scope coverage.

The policy is selected by the trusted local caller. This is not a multi-tenant server: a request’s agent ID is an assertion, not an authenticated credential. A caller that supplies the decision agent’s ID is choosing that synthetic role. Do not expose this command as a public authorization service. The role named scientific does not appoint a scientist, and publication does not grant real publication permission. A new policy file is a new set of assumptions, not independent evidence that the actor deserves those roles.

A principal is the fixture’s capacity/authority key, not automatically a real person or institution. Actual overlapping people across distinct principals, hidden common control and inaccurate qualification declarations remain unmodeled. Internal agent-team expansion alone supplies no new principal or control group.

Version and scope semantics

target.version is a SHA-256 digest of the canonical offer profile: claim ID/text, evidence object, sorted scope labels, resolved dependency targets and required group count. target.evidence_digest separately hashes the evidence object. JSON Boolean, integer and floating-number values remain distinct in the canonical representation. This is the repository adapter’s deterministic JSON convention, not a claim of standards-level canonical JSON or cryptographic signatures.

Consequently, corrected evidence gets a new target version. Correcting a dependency or check contract can also get a new profile version without falsifying a change in evidence bytes. Re-offering the identical profile in one batch is refused. Changing profile metadata does not resolve an earlier reliance record automatically; old and new targets remain distinct. Hashes bind supplied bytes and metadata, not scientific truth or completeness of the declared dependency graph.

Existing check scope cannot be broadened into a new use, duplicate checks cannot multiply coverage, and a reliance cannot lower its offer’s coverage floor. A new offer’s floor is still an author-selected fixture contract, not a proof that the floor is scientifically sufficient. The offering principal alone may amend or change availability for its target under this narrow profile.

Refusals, limits and provenance

With valid CLI arguments, stdout is one JSON object. Exit0 means the batch was processed, including any ordinary per-command refusals. Exit2 means a local input or envelope was invalid. --help and command-line usage errors follow ordinary argument-parser conventions. Requests may come from a named local file or stdin with --request -. No state is persisted between runs.

Inputs are limited to 256KiB each, depth16, 10,000 JSON nodes, 64 commands, 16 principals, 32 agent aliases, 128 array items and 4,096-character strings. Duplicate JSON keys, nonfinite numbers, oversized numeric tokens, unknown schema fields and Boolean values used as integer time/effort are rejected. No panel combination enumeration, arbitrary code, subprocess, remote evidence lookup, credential use or network operation is exposed.

The ledger charges declared check effort per checked scope using the existing runtime. Offer, reliance, amendment, observation, adaptation and administrative labor are not charged. A bounded command count is not a measured compute or human cost guarantee. The separate coupled models study broader participation costs. Currentness uses the supplied simulation clock and local state; a fresh timestamp is not a live evidence observation or delivery receipt.

Each result includes request and trusted-policy digests, accepted/refused command counts, a complete bounded runtime-state summary and a state digest. These make replays comparable; they are not signatures or transferable external approvals. The result’s boundaries object stays machine-readable so an agent can retain these limitations with any derived artifact.

First useful contribution

Run the four supplied cases. Then modify one example to break one claimed property: give two aliases one small capacity, attempt to reuse a stale check, change a scope, hide a dependency, or replace a decision role with a reviewer. Keep the original request, mutation, actual response and a short explanation. A useful contribution is a reproducible counterexample or an adapter that preserves these boundaries, not a new “trust score” without a test.

python3 -m unittest discover -s papers/07-release-packet/agent_api/tests -p 'test_*.py' -v
python3 papers/07-release-packet/agent_api/build_examples.py

Seventeen author tests cover these boundaries; independent launch review lives in ../reviews/agent-launch-red/. The example result manifest records imported runtime and adapter source hashes as well as local source hashes. Submit a proposed patch or issue through the repository’s ordinary owner-reviewed contribution process. This CLI never posts an issue, opens a PR, enrolls anyone or publishes anything.