Add dedicated SI agent candidate; preserve existing plugin

[infotradescout]

Direction and scope

Build a dedicated Selective Intelligence agent instead of continuing the plugin security-repair/resubmission lane. Leave the plugin and canonical free core unchanged. The previous proposed source-hardening patch was not applied.

This draft adds exactly four files under adapters/openai-agent/, based on frozen PR #60 source 3015ede746c84dfb5deac94a622b650701a21e15. It does not merge #60, alter any plugin files, unpublish a listing, install a new SI edition, or create a new hosting service.

Implemented

• OpenAI Managed Agents definition using the canonical SI skill and its existing Orchestrator/Worker/Objector roles, web search, and configured separate subagents.
• Isolated hosted-sandbox bootstrap pinned to exact source commit and tree, with actual file-byte verification, modified/symlinked-source rejection, anonymous Git, and no downloaded installer execution.
• Offline configuration export and a registration-only helper with fixed-origin requests, explicit project binding, preserved creation intent, returned-ID receipts, readback verification, and no blind retries after uncertain responses.
• Task-session request preparation only; this helper cannot start a session, run inference, publish an agent, or delete an agent.

Executed evidence

On Python 3.13.5, python -B -m unittest -v test_agent passed 19/19 tests with zero failures. Scope: offline configuration/registration controls and real local synthetic Git fixtures. Provider replies in registration tests are synthetic. No model, hosted sandbox, or production-tool acceptance is inferred.

GitHub readback at 46be3f6fa78fdb7e76b99c9f7040a836615e7cc2 matched all four locally tested blobs:

• README.md: 52afeca94d3cd5b3fdf3c7d4b872f09a263bfa70
• agent.py: a8015ee7f8f8f8838a6e83d20f0264d40306df69
• bootstrap.py: b8ccb9323f8055e71e73ab1e46eeead46887e6ea
• test_agent.py: 01858c75e12f7f92955d252dc5ee0d550badf6a0

Source comparison confirms only these four additions and zero deletions. The commit uses [skip ci]; no hosted CI success is claimed. Review was same-context/degraded, not independent.

Open boundaries

No authorized OpenAI application credential/project binding or authenticated connected browser was available. Zero OpenAI API calls, registrations, billable sessions, or live tasks were performed. The agent does not yet exist in the owner's OpenAI project.

The full pinned repository could not be cloned in the execution container because network/DNS access was unavailable. Exact source identity and governing files were retrieved through the connected GitHub app. Full pinned-source bootstrap, full repository/project-index refresh, hosted prerequisites, live subagent behavior, task completion, and checkpoint recovery remain unverified.

Operational connectors and Infinity are not fabricated or inherited from ChatGPT. No new spending authorization is assumed; the free portable core remains unchanged.

Next exact action

Register the saved definition in the owner-designated OpenAI project using authorized credentials outside chat/sandbox, retain and read back the provider ID, then validate the pinned bootstrap and one bounded real task under an approved execution budget. Prove actual session continuation before claiming recovery. Keep this PR draft until that acceptance boundary is satisfied.

[infotradescout]

Master continuation — registration identity repair, 2026-10-02

Published head: 72264189b82353f9ac58ade4442e64fc62b05028 (two-file follow-up to 46be3f6). This PR remains draft and unmerged. The canonical plugin, free core, bootstrap, source pin and agent definition are unchanged.

Reproduced offline on exact original bytes: POST returned agent_created, GET returned the same definition under agent_wrong, and the helper incorrectly wrote registered_readback_verified_not_live_tested. The fix requires readback ID equality before a verified receipt, rejects non-object creation/readback and malformed tool entries with explicit errors, and retains intent/created-ID receipts so a failed check cannot recreate the agent blindly.

Executed on Linux Python 3.13.5, using actual helper code and synthetic request callbacks (no provider):

• Original source/original suite: 19/19 pass.
• Expanded unchanged tests against original source: 27 tests, 9 failed subcases and 12 error subcases, retained as the negative control.
• Identical expanded tests against repaired source: 27/27 pass, zero failures/errors/skips.
• Original AgentTests and SourceTests ASTs, including all their assertions, are unchanged.
• Source and test originals matched Git blobs before execution; returned remote content hashes match tested repaired blobs: agent.py 532e7a436861dc93f2c838b4f9a6a0820f5ebb03; test_agent.py 5954d1f25af2d3a6b1801a14855f543ae7baf832. GitHub compare confirms only those two files changed.

Tested scope includes missing/wrong/invalid readback identities, non-object responses, malformed tool entries, normal provider metadata, saved created ID after readback timeout, and refusal to repeat creation. This is bounded offline behavior, not evidence that OpenAI ever returned these malformed replies.

Public network/DNS failed in the execution container, so exact files were retrieved through the connected GitHub reader and reconstructed/hash-verified for local execution. Full repository clone/project-index refresh, full repository gate, independent-context review, hosted bootstrap, real registration/session/continuation and operational connector acceptance remain open. Review in this continuation is same-context, not independent. Commits use [skip ci]; no hosted CI pass is claimed. Zero OpenAI API calls, billable sessions, model requests, production writes or new services occurred.

Next: use the existing authorized project/credential binding when available to register/read back this repaired definition, preserving all original intent receipts and the separate hosted/budget acceptance boundary. Do not reopen plugin repair or repeat the identity investigation without a relevant change.