You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Wardnet must preserve fail-closed credential/admission, framing, header-minimization and streaming security behavior at its LLM-facing gateway seam without becoming the model/provider-routing authority.
contextual-orchestrator (CO) is the canonical owner of production LLM provider/model discovery, routing/fallback, concrete model selection, provider credentials and provider-specific transport differences. Wardnet must not own a LiteLLM deployment surface, virtual-key authority, provider/model/group selector, provider credential registry, or direct paid-provider fallback.
Wardnet remains responsible for the gateway/security-control side of the seam: fail-closed admission of the request/credential class required by a released CO contract, bounded request/framing validation, header minimization, security evidence, credential non-disclosure, safe streaming relay where Wardnet actually proxies a released CO interface, and denial before upstream work when Wardnet policy fails.
Current preservation lane — 2026-09-05
PR #88 remains open at exact 98a935e4b058d0500d520425b8ebf6ff4106aa1b and GitHub now correctly records it as Draft. It is non-mergeable by architecture in its current direct-LiteLLM shape and remains a preservation lane rather than the production integration vehicle.
Fresh contextual-orchestrator GitHub Release inventory is still empty. Wardnet therefore has no immutable compatible CO API/client/schema Agent to pin. A mutable CO branch/head, sibling checkout, direct provider SDK, or copied routing implementation is not an acceptable substitute.
Do not close #88 merely because its architecture is superseded. Preserve and later transfer every still-valid Wardnet-owned delta, test, fixture, security rationale and hostile proof before a successor retires it.
Evidence that must survive reconstruction
bounded Authorization/request credential grammar and fail-closed missing/duplicate/wrong-scheme/malformed/non-ASCII/oversized handling;
zero-upstream-hit negative tests for invalid credential classes;
no submitted credential or masked fragment in errors/events/audit/support/telemetry;
narrow request-header forwarding with cookies, management credentials, forwarding/proxy, host, framing and unrelated trace baggage stripped;
redirects and ambient proxies disabled where Wardnet owns the HTTP boundary;
request path/query/method and bounded-body preservation required by the released upstream contract;
streaming response relay without whole-response buffering;
safe correlation/rate-limit metadata only when the released upstream contract permits it;
stable property tests and fuzz coverage for the same untrusted credential/request grammar;
Historical LiteLLM/provider-specific implementation details are preservation evidence only. They do not become current architecture authority.
Replacement contract after CO release
When contextual-orchestrator publishes an immutable compatible release, reconstruct the Wardnet consumer slice from fresh protected Wardnet main through an Anti-Corruption Layer.
Consume only the released CO API/client/schema Agent and bind exact version/digest/provenance.
Remove direct LiteLLM/provider/model/group/paid-fallback authority from Wardnet.
Keep model/provider selection absent from Wardnet-owned domain/API state unless CO exposes an opaque capability/request field that grants no provider authority.
Fail closed when the required CO release/schema/capability/provenance is absent, stale, malformed, incompatible or unverifiable.
Transfer the Wardnet-owned security regressions above against the actual released CO boundary and prove zero direct provider/LiteLLM calls from Wardnet.
Keep GitHub Actions model-backed use on the organization-owned orchestrator/free gateway contract only; no provider/model/group/paid fallback pin belongs in Wardnet.
Update the Wardnet Context Map/ADR/SECURITY/THREAT_MODEL/TRACEABILITY/CHANGELOG and the sole product-gap writer with released-contract evidence rather than mutable source claims.
Wardnet does not authenticate provider virtual-key existence, revocation, budget, team or model scope.
Wardnet does not manufacture or repair a provider credential.
Wardnet does not choose a concrete model/provider or silently fall back to a paid provider.
Wardnet does not embed contextual-orchestrator or copy its routing implementation.
This issue does not weaken non-LLM gateway authentication or generic request-security controls.
Completion gate
Close only after an immutable compatible CO release exists and a protected-main Wardnet successor consumes it through a versioned ACL with every still-valid #88 Wardnet-owned security delta demonstrably transferred or specifically proven obsolete. The successor must have one unchanged exact head with terminal repository/security/coverage/package/SBOM/provenance/review/thread/governance evidence. No mutable foreign dependency, self/model approval, routine bypass, force push, gate weakening or predecessor-evidence reuse.
Problem and authority boundary
Wardnet must preserve fail-closed credential/admission, framing, header-minimization and streaming security behavior at its LLM-facing gateway seam without becoming the model/provider-routing authority.
contextual-orchestrator(CO) is the canonical owner of production LLM provider/model discovery, routing/fallback, concrete model selection, provider credentials and provider-specific transport differences. Wardnet must not own a LiteLLM deployment surface, virtual-key authority, provider/model/group selector, provider credential registry, or direct paid-provider fallback.Wardnet remains responsible for the gateway/security-control side of the seam: fail-closed admission of the request/credential class required by a released CO contract, bounded request/framing validation, header minimization, security evidence, credential non-disclosure, safe streaming relay where Wardnet actually proxies a released CO interface, and denial before upstream work when Wardnet policy fails.
Current preservation lane — 2026-09-05
PR #88 remains open at exact
98a935e4b058d0500d520425b8ebf6ff4106aa1band GitHub now correctly records it as Draft. It is non-mergeable by architecture in its current direct-LiteLLM shape and remains a preservation lane rather than the production integration vehicle.Fresh
contextual-orchestratorGitHub Release inventory is still empty. Wardnet therefore has no immutable compatible CO API/client/schema Agent to pin. A mutable CO branch/head, sibling checkout, direct provider SDK, or copied routing implementation is not an acceptable substitute.Do not close #88 merely because its architecture is superseded. Preserve and later transfer every still-valid Wardnet-owned delta, test, fixture, security rationale and hostile proof before a successor retires it.
Evidence that must survive reconstruction
Historical LiteLLM/provider-specific implementation details are preservation evidence only. They do not become current architecture authority.
Replacement contract after CO release
When contextual-orchestrator publishes an immutable compatible release, reconstruct the Wardnet consumer slice from fresh protected Wardnet
mainthrough an Anti-Corruption Layer.orchestrator/freegateway contract only; no provider/model/group/paid fallback pin belongs in Wardnet.Explicit non-goals
Completion gate
Close only after an immutable compatible CO release exists and a protected-main Wardnet successor consumes it through a versioned ACL with every still-valid #88 Wardnet-owned security delta demonstrably transferred or specifically proven obsolete. The successor must have one unchanged exact head with terminal repository/security/coverage/package/SBOM/provenance/review/thread/governance evidence. No mutable foreign dependency, self/model approval, routine bypass, force push, gate weakening or predecessor-evidence reuse.