1.7: package descriptor schema draft + 3 validated descriptors (phase 1 unit 1.7)

[cunninghamcard-bit]

Phase 1 unit 1.7: descriptor schema draft. It follows arcships#200 §6.5 (docs/aisdk-architecture-alignment.md:1324) and RFC-0040 §2.1 (rfc/0040-provider-auth.md, from commit 3757cfd; that file is not in exp/restart). The branch is cut from exp/restart @ 8455db9, and the PR base is exp/restart.

Added

• descriptors/package-descriptor.schema.json (JSON Schema 2020-12).

  • It requires all 16 descriptor fields listed in docs: full-chain AI SDK alignment design and impact map arcships/aimux#200 §6.5. In the schema they are named package_id, factory, settings_schema, credential_fields, runtime_handle_fields, methods (holding the method list and default_method), name_semantics, provider_string_rules, provider_options_rules, tools, extensions (holding discovery), auth (holding offline_auth_recipe), catalogue_mapping, audit_metadata, supported_features and source_verification.
  • additionalProperties: false is set throughout.
  • It encodes these rules from RFC-0040 §2.1:
    • A descriptor carries no credentials.
    • Credential fields are the only ones that accept {"env": ...}.
    • Runtime handles (fetch, id generator, ws) are listed apart from the settings data, so a closure is never treated as data.
    • The protocol/family tag is for audit and test grouping only.

• descriptors/{openai,anthropic,openai-compatible}.json, written from the Rust factories and settings:

  • openai/mod.rs:187, :239, :309-385
  • anthropic/mod.rs:110, :164, :255-295
  • openai_compatible/mod.rs:60, :120, :208-245

  Each pins its SDK version from fixtures/aisdk/VERSIONS.json (openai 4.0.83, anthropic 4.0.71, openai-compatible 3.0.62). Tool ids are the ones found in the source.

• aimux-providers/tests/descriptor_schema_test.rs, plus a jsonschema dev-dependency (a workspace dependency already present for aimux-core). It has four tests:

  1. every_descriptor_validates: all 3 descriptors pass the schema. This is the PHASES acceptance check.
  2. schema_rejects_malformed_descriptors: the schema rejects each of 8 edits, so it is not vacuous:
     • a missing required field
     • an unknown top-level field (body_overrides)
     • settings that allow unknown keys
     • a badly formed package id
     • an unknown method kind
     • an unknown auth mode
     • a closure listed as a runtime handle
     • a non-semver SDK version
  3. descriptors_are_internally_consistent: checks rules the schema can't express:
     • credential fields are names in the settings schema
     • runtime handles are not listed as settings data
     • the default method is in the method list
     • every provider-string example names a listed method
  4. provider_string_examples_match_the_factories: each descriptor's provider-string examples equal what create_openai, create_anthropic and create_openai_compatible actually report. The callable provider's default method matches default_method.

The descriptors are not wired into anything yet. Generation and the manifest come later (1.8 and phase 3). Fixture and cassette diff: zero.

Results (local linux x64)

	R.0 baseline	1.7
cargo test --workspace	2871 / 0 / 7	2875 / 0 / 7 (+4)
fmt / clippy -D warnings	ok	ok
boundary script	0	0
gen_ts_types --check	ok	ok
Node ava (node 22)	87 / 1 fail	87 passed, same 1 failure

Python was not run (no venv). The schema field names and shapes are a draft for review. arcships#200 §6.5 names the fields but not their structure. "arcships#214 factories" and "arcships#219 types" in PHASES refer to upstream PRs; the samples here were written from the current code that came out of them.