Changelog

Unreleased

• Test insufficient low effort and sufficient high effort within Grok's maintained task role.

• Establish Grok role suitability and high-effort judgment margin in the positive current-version selection fixture.

• Grade unavailable routes on refusal and owner control; explicitly elicit fallback consent actions, user-owned task prerequisites, and leader responsibility assignments.

• Require independent catalog and executor proof for every fallback pair, and exercise unsuitable and suitable Cursor reviews, useful Ultra without panels, witnessed-route prerequisites, and concrete specialist input selection.

• Clarify the discovery boundary between delegation and model selection in the delegation skill's description.

• Route ordinary demanding work to GPT-6.1 Sol or Opus, prefer proven newer same-name GPT/Claude versions, restore Sonnet 5.5 for its shallow-cognition middle tier, and require medium-or-higher effort for Astra/Fable and Sol/Opus. Bind OpenAI/Anthropic max to all model, frontier-problem, panel-loop, and capability gates; encourage useful Ultra/Ultracode panels with independently selected participant efforts.

• Separate ordinary same-leader invocation plans and live selection records from explicitly witnessed receipts, preserving native assurance limits and each consumer's existing authority and verification gates.

• Re-prove the original account binding before every task continuation, bind alternate-account Codex CLI resumes with explicit CODEX_HOME, and reject ambient resumes before sending task data when account affinity is unproven.

• Define target, relationship, ownership, transport, and assurance as independent delegation dimensions, with model choice kept separate.

• Require controller-observed evidence for every fact an ordinary consumer relies on and product-attested ChatGPT Codex child evidence for canonical publication; self-reported dimensions remain diagnostic and non-gating.

• Bind the consumer's per-dimension assurance minimum before dispatch and reject any observed target, model, topology, authority, or result evidence below it.

• Publish the qualification order for separate ChatGPT Codex and Codex CLI/TUI surfaces, followed by Claude Code and Claude Desktop, then Cursor and Cursor Agent, without claiming that validator-only support can issue evidence.

• Register the strict generic dispatch-evidence owner with Task Witness and emit a canonical projection without taking review or publication authority.

• Bind producer and issuer identities to a separate native adapter that deterministically freezes already-observed execution facts without claiming to launch or authenticate the worker itself.

• Keep the native adapter issuer contract independent from the dispatch- evidence bundle contract.

• Fail closed by omitting the unauthenticated producer and issuer from the current provider; retain only the validator and its bootstrap-renderer module until an authenticated native harness integration exists.

• Preserve failed, timed-out, incomplete, and other nonclean execution evidence by accepting strict Boolean usable: false and projecting the exact status.

• Add separate ChatGPT Codex and Codex CLI/TUI native-child profiles with an executable pre-spawn freeze and post-result recording seam. Bind exact plan/request identity and native launch, status, result, agent, session, and context observations without parsing worker text as host evidence.

• Keep native model and effective-authority assurance self-reported until the product exposes an exact model and per-spawn enforcement observation. Reject stronger consumer minima before spawn and keep the binding outside provider producer and issuer registration.

• Separate native transport completion from verified usability. Bind the same-leader verification observation and strict usability Boolean, preserve completed-but-blocked or unverified results as unusable, and reject a usable non-completed terminal.

1.0.0

Initial Agent Plugins v1 package for delegation topology, bounded authority, live-capability model selection, and leader integration. The root manifest owns the 1.0.0 identity, Codex metadata uses extensions.com.openai.interface, and Claude Code uses a validated native projection.