Provider agent contracts research corpus
This corpus is the research input behind the platform's credential plane: a per-provider study of how the commercial agent platforms design three things, sessions, agents, and vault secrets, studied independently of each other and then read back against our contract. It follows the same method as the agent platform corpus, narrowed to the axis that corpus covers least: where a third-party credential lives, what it binds to, and who holds it at the moment it is used. Where a conclusion here differs from an accepted record in the ADR index, the ADR is authoritative.
Method
The research prompt is preserved so the scope and evidence rules behind every dossier remain reproducible. Each provider was researched in isolation, with no provider's findings visible to another provider's run, so that agreement between two dossiers is evidence of convergence rather than an artifact of the order they were written in.
Product dossiers
- Anthropic Claude Managed Agents
- OpenAI Agents API
- OpenComputer Serverless Agents
- xAI platform
- LiteLLM proxy, a gateway that normalises many providers behind one contract rather than a provider itself. It was added after the combined proposal, against an evidence snapshot dated 2026-10-13, and closes with what its findings would change or confirm in that proposal.
Three of the four agent providers also have dossiers in the agent platform corpus, written against a different question and frozen before this study ran: Claude Managed Agents, OpenAI Agents API, and OpenComputer. Those remain the record of what the agent-platform study concluded at the time. This corpus is the deeper read on sessions and credentials, and it supersedes neither. xAI has no agent-platform dossier; its client-side coding agent is covered in the ACP corpus instead, and this corpus records why: there is no server-hosted agent product to write an agent-platform dossier about.
Combined proposal
- Combined schema proposal, the four dossiers read together against
proto/trogonai/agents/agents/v1alpha1andproto/trogonai/session/sessions/v1alpha1. Analysis and proposal only; every change it argues for is gated on an ADR.
Status
The four provider dossiers are complete for the evidence snapshot dated 2026-09-11; the LiteLLM dossier is a later 2026-10-13 snapshot. The combined proposal is a draft proposal, not a decision: it closes some questions that were open (egress substitution now has four independent implementations, so it is no longer a live design choice) and opens others that each want their own record.
One finding has since been recorded. Draft ADR#0063 takes the declaration half of the proposal's §6.1 and adds ConnectionDeclaration to AgentDependencies, citing this corpus for the two findings that fix its shape: that material rotates while structural identity does not, which is why that declaration is the one dependency in the contract that pins no version, and that the secret never enters the agent process. Everything else here remains a draft proposal rather than a decision, and the proposal's sequencing records what the first pass through the contract changed about the order of the rest.
The headline findings, in the order they carry weight. Each is scoped to the providers whose documentation supports it, because the weight of a finding in this corpus is the number of independent designs that reached it, and a bullet that reads as a universal rule destroys exactly that information:
- Material rotates and structural identity does not. Anthropic and OpenAI reached the same rule with no shared specification between them: the secret behind a credential may be replaced without changing the credential's id, authentication type, or bound server URL, while those three may not change at all.
- Write-only is expressed by schema shape, not by a redaction rule. OpenAI. The create type carries the secret and the read type simply has no field for it, so there is nothing to redact and no placeholder to leak through. Anthropic and OpenComputer are also write-only but do not document the property as a consequence of the type split.
- Credentials attach to the session, not to the agent definition. Anthropic (
vault_idson session create) and OpenAI (session-level vault attachment). This is the one place a pin-at-admission contract has to make an exception, because a pinned credential cannot be revoked. Note what it does not buy: OpenAI also permits acredential_idselector on an agent tool, and deleting a stored credential there revokes neither the provider-side token nor a session already running on it. - The secret never enters the agent process. Anthropic and OpenComputer, plus the two open-source implementations already on file. Substitution happens at network egress from an opaque placeholder, and it is a set plus a strip, not a set alone. Scheme enforcement, redirect revalidation, and response isolation are undocumented in all four, so the mechanism as published bounds where a credential reaches rather than what a compromised agent can learn by using it.
- The negative result is explanatory. xAI has no vault and no provider-held MCP credential flow, and that is not a gap in xAI's design: a client-side agent holding the user's own credentials needs no server-side vault. The vault is a consequence of hosting someone else's agent against someone else's credentials. xAI is not credential-free, though, and the dossier inventories what it does have: inference API keys, management keys, ephemeral realtime tokens, mTLS certificates, an undocumented OAuth bearer path, and CLI credentials on the user's own disk.
- One finding is single-source and is labeled as such. Archive purges the payload and keeps the record while delete keeps nothing is Anthropic only. The proposal argues for reusing our existing erasure idiom on that basis, which is an argument from our own contract rather than from convergence.
