Skip to content

Make Codex and Claude Code subscriptions go further: a local model router that gives routine agent work to economical models and keeps your strongest model for the hard parts.

34.5 KB
Raw Blame

Connecting and managing models

Purpose

A saved connection owns a stable source identity, protocol endpoints, credentials, and selected model bindings. Templates and metadata assist configuration; saved source membership and current runtime authorization determine actual access.

Requirements

Requirement: The API wizard uses the current executable directory

Desktop MUST present service selection, connection configuration, and model selection as separate steps. Provider/product counts, search, billing filters, and branding come from the complete current executable catalog rather than a fixed sample. A provider with several products requires an explicit product choice; custom, reconnect, and dedicated free-resource entry points may start at the appropriate configuration step without expanding their catalog scope.

Every connection has an editable display name, initially the chosen product name. The name is a label, not supplier/account/endpoint/model identity. New Desktop names must pass its nonempty, 60-character, and duplicate-name checks; existing same-name sources remain readable and independent. Check, model selection, save, and recheck preserve the user's valid name without replacing it with a template.

Back/forward within the wizard preserves each product's separate non-secret draft and current-session credential input. Switching products cannot copy a Key into another product. Closing, discarding, or saving releases temporary credentials; ordinary persistence never stores them. Editing check-relevant facts invalidates the previous check. Internal navigation is not abandonment; close uses the normal unsaved-edit rule. None of these actions implicitly starts inference.

Requirement: Connection identity survives supported lifecycle edits

One template may create multiple independent connections even with identical provider/model names or Key strings. Source and model views retain the connection label and separate product identity. Rename, recheck, and append retain source and existing Binding identity. Append adds only explicitly selected new models, without duplicates, replacing credentials, disabling existing models, or changing old facts after a directory refresh. Refresh retains the currently selected source.

Desktop and CLI MUST share rename, append, remove-model, delete-source, endpoint, Key, and enabled-state semantics. Removing one model preserves remaining members and shared credentials. Removing the last model requires explicit whole-source deletion; removing the last Key instead leaves a repairable source. Unreachable sources may be saved inactive, and failed replacement cannot overwrite a working configuration with partially prepared state.

Scenario: Recheck reveals additional models

  • WHEN a saved registered source's current account directory contains models not yet imported
  • THEN they MUST be selectable for explicit append, while existing members, Keys, enabled state, label, and bindings remain stable.

Requirement: Deletion respects exact references and owned secrets

Deleting or narrowing membership MUST inspect exact Bindings, including active and disabled retained routes, held execution/Continue versions, drafts, and unreconciled recovery references. Unknown or failed reference reads block deletion. Concurrent references are rechecked inside the writer; previews are not authority. UI shows the blocking objects and does not silently rewrite their plans. Historical usage is preserved.

An explicit publication checkpoint MAY rotate old recovery references only from the verified installed active publication. Preview binds its exact digest and workspace revisions; Apply rechecks under the writer and advances publication revision once through normal prepare/install/activate/rollback/restart. It preserves plans, heads/tombstones, grants, aliases, and permission facts, replacing the old last-known-good only after success. Empty active state remains NoNewCalls. The checkpoint neither deletes a source nor invokes inference; unresolved/failed checkpoint is not permission to delete or manually edit recovery snapshots.

Source deletion MUST atomically manage the source and its HiRoute-owned secrets through a recoverable Operation. Compensation and retry preserve exact secret/source identity and cannot overwrite later configuration. Subscription removal preserves native login/configuration; late checks, saves, and maintenance cannot recreate the removed source. A future scan requires explicit approval again. Released legacy save semantics remain replacement semantics; append/delete are explicit intents.

Requirement: Raw upstream model IDs are opaque, not internal identifiers

CLI/API, persistence, compilation, publication, and credential authorization MUST preserve an upstream ID exactly: nonempty, at most 512 UTF-8 bytes, no leading or trailing Unicode White_Space, and no Unicode Cc controls. Unicode, punctuation, and internal spaces are legal; U+FEFF is preserved because it is not White_Space. No lowercasing, transliteration, normalization, or escaping may replace its identity. Internal source/Binding/offer references retain their stricter identifier rules.

Desktop's manual entry policy currently accepts only ASCII for new or changed raw IDs, with the same whitespace/control/byte bounds. Non-ASCII display names remain separate. Existing, directory-selected, and metadata-prefilled non-ASCII IDs retain their exact value when other fields are edited; replacing one with a different non-ASCII manual value fails before check, inference, or save. A malformed legacy model cannot break the entire management snapshot: retain readable state and report that model's execution restriction.

Scenario: Rename a model whose existing ID is Unicode

  • WHEN the user edits only its display name or capabilities
  • THEN the original ID MUST survive unchanged and be sent exactly at execution; the manual ASCII policy cannot force a replacement.

Requirement: Registered endpoint facts remain server-owned

Executable registered options MUST resolve from the client-version-bound ReleaseFacts registry and the selected validated official model snapshot, with embedded model data as the documented offline fallback. Manifest identity, digests, cross-references and supported schemas SHALL be checked before use. An arbitrary data-directory catalog, an unvalidated held copy or a schema suffix SHALL not become a selectable source of execution authority.

Connectors, endpoint/protocol profiles, authentication, secret slots, static headers, adapters, connection options and billing classes SHALL remain owned by the installed registry. A model binding becomes executable only when its exact references and required adapter are implemented, its offer matches the endpoint's entitlement/billing scope, and the existing source/account checks permit use. Descriptive-only models remain queryable. A known minimum compatible client SHALL be shown when available; otherwise the missing adapter or qualification SHALL be stated without inventing a version.

A registered check SHALL submit option identity, expected provenance, edit correlation, selected models and protected input. The daemon SHALL resolve exact destination, protocol, authentication, headers, identity and billing and reject client-supplied overrides of registered authority. Native API uses direct HiRoute adapters. Public endpoint hints may prefill a new editable connection, but a metadata refresh SHALL NOT rewrite a saved destination or expand Key permissions.

Saved source/binding identity, selected membership, credentials, target and historical admission provenance SHALL survive compatible metadata updates. A catalog-derived capability SHALL resolve from the current selected snapshot for new authoring/publication. A catalog digest change or a compatible change of context limit, tool/vision/streaming support or supported reasoning values alone SHALL NOT invalidate an otherwise legal saved connection. Existing user-declared values SHALL retain their source-local meaning and editable authority. Refresh SHALL neither fabricate a successful connection check nor overwrite historical save evidence.

New selection, check, recheck and publication SHALL qualify against one current snapshot. A withdrawn model, unsupported adapter, invalid source/account, or previously selected reasoning setting that is no longer supported SHALL produce a specific unavailable/conflict reason. The user SHALL be able to review a valid replacement; the service SHALL NOT silently change reasoning choice, drop selected members or publish a replacement Plan. Renaming a source, removing another member or changing Keys SHALL preserve unchanged saved model values and identity; newly added/rechecked members receive current admission. Withdrawal SHALL not prevent an otherwise legal removal/disable action.

Already published revisions SHALL keep their compiled model facts, candidate order and reasoning configuration. Their existing runs, new sessions, newly delegated Workers and Continue SHALL not depend on the mutable catalog or held copy, including after restart. Current authorization, cancellation and credential revocation still apply. Applying new capability defaults to execution SHALL use normal preview/publication. New logical requests through old publications still obtain current source-scoped valuation independently of frozen execution; settled history remains unchanged.

When a later client includes downloaded facts, the selected data revision and actual connection compatibility SHALL govern reuse; embedded presence alone SHALL not suppress newer official data or force a new connection check. A saved member that has no usable current official facts after withdrawal or cache removal SHALL report the actual new-use limitation, without erasing its identity or frozen publications.

Scenario: Existing registered model receives new limits

  • WHEN a compatible official update changes a saved model's catalog-derived context limit while endpoint, account and adapter remain valid
  • THEN the model remains selectable with the new default for authoring; source/binding IDs, Keys and membership are preserved and no connection check is claimed.

Scenario: Preserve manual and ambiguous capabilities

  • WHEN a saved custom/native declaration has explicit or indistinguishable retained user values while catalog defaults change
  • THEN those values remain intact; catalog refresh does not convert them into registered authority or discard the connection.

Scenario: Previously selected reasoning is removed

  • WHEN a refreshed supported-reasoning set excludes the value selected by a new draft
  • THEN new publication explains the unsupported selection and requires an explicit valid choice; the existing published revision retains its prior setting.

Scenario: New model uses an installed service

  • WHEN the official catalog adds a model for an installed endpoint and the normal account/connection check succeeds
  • THEN it saves and qualifies under the ordinary source lifecycle with the supplying snapshot recorded as historical evidence.

Scenario: Metadata is readable before execution support

  • WHEN a model record requires an adapter the client does not implement
  • THEN its metadata remains readable while executable selection/check/publication is unavailable with a specific reason and a minimum version only when established.

Scenario: Withdrawal and restart preserve published execution

  • WHEN an official model is withdrawn or its only usable downloaded facts are removed, and the service restarts
  • THEN new use requiring the absent facts is unavailable with a reason, while existing publications, new sessions through them and Worker Start/Continue use their frozen execution facts.

Scenario: Routine lifecycle edits preserve retained members

  • WHEN the user renames a source, removes a different model or changes Keys after catalog capabilities changed
  • THEN unchanged saved members keep their values and identities; only newly added/rechecked members undergo current admission, and removing or disabling an unavailable member remains possible.

Scenario: Upgrade absorbs a downloaded model

  • WHEN a later client embeds a model already selected in a saved connection
  • THEN the compatible connection continues without a forced recheck, and comparable data revisions choose defaults without replacing a newer usable cache.

Scenario: Public connection hint changes

  • WHEN a publication changes a provider's suggested Base URL or authentication description
  • THEN the new hint can appear in editable setup, but existing saved destinations, protected credentials and registered authority are unchanged.

Source SPEC comments:

Requirement: Directory access and inference evidence are independent

Model directory reads are optional enhancements. They use the selected account, retain valid items amid malformed entries, and distinguish unsupported, empty, invalid, authentication, permission, quota, timeout, and connectivity outcomes. A registered or manually declared supported endpoint can be saved without per-model endpoint records, a directory, or inference probing. An unrecognized ID retains raw/provider/source scope with no borrowed canonical configuration, score, price, or free entitlement. Safe text fallback or user declarations are labeled untested.

Check/refresh is explicit and bounded, does not import automatically, perform paid inference, cycle through unselected targets, or establish continuous health. Cancellation stops further probe requests; late results cannot replace later edits. A real model/tool test requires explicit invocation and proves only what it tested. Protected save retains candidate correlation, version checks, and recoverable result handling even if a check succeeded earlier.

Scenario: Directory is missing but the endpoint is configured correctly

  • WHEN the account offers no directory or the directory operation fails
  • THEN the user MUST be able to enter models/capabilities manually, preserve unknown fields, and save supported configuration without a synthetic inference.

Requirement: Custom connections preserve protocol and transport scope

A Native API source may have at most one endpoint per Responses, Messages, and Chat Completions protocol, sharing model selection and one Key pool. All endpoints are unauthenticated or use that pool, though authenticated header forms may differ. Each endpoint owns its full target, auth, and necessary static headers. Same-protocol execution is preferred; an absent protocol may use only an expressible existing conversion. Endpoint health is isolated; manual Key disable affects the source.

Custom configuration supports name, base URL/path prefix, None/Bearer/specified Header authentication, and multiple manual models. A user may explicitly configure HTTP or HTTPS for a loopback, LAN or remote endpoint, whether its host is a literal address or a hostname. HTTP is not rejected solely because the destination is non-loopback. HTTPS is recommended; an HTTP selection explains that API keys and request/response content travel in plaintext without adding a mandatory confirmation. Target normalization preserves the configured hostname; DNS and connectivity are evaluated by bounded request transport, not by passive discovery or a syntactic check. The same transport policy applies to durable source validation, publication and actual Gateway execution. Registered catalog destinations remain bound to their exact HTTPS authority; managed CPA transport retains its numeric-loopback scope. Target changes require revalidation. Credentials cannot be redirected to an unauthorized destination; TLS failure never falls back to plain transport. Embedded URL passwords/query Keys must use protected input instead.

Vision, tools, context, output, streaming, native reasoning, and price may remain unknown. Trusted facts, probe evidence, user declaration, and unknown are distinct. Contradictory parameters or unrepresentable reasoning are rejected rather than silently stripped. Missing canonical match, scores, or price alone cannot block an otherwise expressible model. Cost hints without currency/unit/product/region/time semantics are not PriceRate, free entitlement, or savings evidence. Metadata field closure and reference pricing follow model facts.

Scenario: Save an explicitly configured HTTP source without a directory

  • WHEN the owner configures a valid LAN or remote HTTP endpoint, leaves the model directory path empty and supplies supported manual model declarations
  • THEN connection check, save and saved-source recheck accept that target without DNS or directory/inference calls, preserve the exact target and authentication, and do not claim reachability, account access or inference has been verified.

Scenario: Correct an invalid connection target

  • WHEN a connection check rejects an invalid URL, unsafe path, authentication header or protocol header, including an additional protocol endpoint
  • THEN the daemon/API retains a closed, specific reason in the existing error envelope, with the released INVALID_ARGUMENTS machine code; Desktop presents a localized correction and retains input. Errors never expose URL credentials, protected keys or raw provider responses. Other URL/path/authentication checks, redirect restrictions and HTTPS certificate validation remain enforced.

Source Issue: https://github.com/higress-group/HiRoute/issues/154

Requirement: Saved-source recheck reconstructs protected facts

Recheck submits source ID, exact expected revision, permitted candidate reference, edit revision, and check ID. It MUST NOT accept caller-provided lineage, endpoint, option, protocol, credential identity/value, or capability facts. Daemon reconstructs registered sources from current catalog plus saved identity, and custom sources from their private safe recheck descriptor. A legacy custom source lacking that descriptor fails closed; connector-owned subscriptions use their reauthorization workflow. Recheck shares deduplication, cancellation, candidate, save, and recovery contracts and reveals no reconstruction secrets.

Native discovery/preparation/preview associations are process-local, not durable permissions. The trusted process retains full effective configuration layers, scanner/source revisions, credential descriptors, and catalog evidence. It rechecks both complete evidence and protected credential fingerprint before Preview and, after idempotency/version checks, before writer admission begins the Operation. A changed high-priority endpoint/model invalidates a candidate even if the token file is unchanged. Exact prepare_id replay returns the same candidate; reuse for a different discovery conflicts. Prepare is synchronous with no cancellation endpoint: closing drops UI continuation, and late results are discarded by edit correlation. Restart requires new Scan/Prepare; admitted Apply/Result remains durable.

Executable resolution requires a regular executable and bounded process/version handling, not a new owner/group-writable-parent trust rule. Discovery cannot run native configuration helpers/hooks. Missing managed Codex runtime is runtime_unavailable, not a successful empty scan; an available runtime without native login may legitimately return a complete empty result.

Requirement: Subscription modes retain their credential owner

Codex and Claude share check/select/save/disable/cancel/recovery workflows with provider-specific readers. Each provider has at most one enabled account/mode. Independent CPA sign-in is the default recommendation; NativeBorrowed is an explicit option and saved choices do not migrate automatically. Login completion alone does not enable routes or import all models.

Desktop subscription checks and model selection SHALL use subscription-specific titles. Native discovery controls, including Hermes profile scanning, SHALL remain on the device scan list and stay hidden in subscription details. Returning to the scan list restores its controls; cancelling proxy editing retains the selected subscription context.

A successful subscription save ends its Desktop check interaction. A later explicit check, including another provider or account, MUST proceed without releasing or invalidating the saved access. Running checks and verified checks that have not completed saving retain their existing interaction ownership.

Mode Credential authority Repair action
NativeBorrowed Native client owns login/refresh; CPA receives current access leases only Update native login and explicitly recheck
CpaManaged CPA owns an isolated private auth directory and independent OAuth refresh Repeat independent sign-in as required

Stable provider account identity, not email, filename, or auth-index, determines matching. A same-account mode switch requires check/save and retains selected source and unremoved models. Another account requires explicit check/selection. Pending login is isolated from the enabled instance; saving reuses its auth directory with one refresh writer. OAuth callback binds exact session/state/provider through protected input, not logs or ordinary Operations. Cancel, expiry, restart maintenance, and removal fence late callbacks and unsaved credentials. Disable revokes admission and stops refresh while retaining credentials; remove-login first revokes saved route access, then stops the exact instance and deletes its owned credentials. Neither claims provider-global OAuth revocation.

Requirement: Independent sign-in receives browser callbacks locally

CPA owns temporary browser callback listeners for the existing Codex http://localhost:1455/auth/callback and Claude http://localhost:54545/callback redirects. Bind only available IPv4/IPv6 loopback families. Each provider has at most one pending sign-in; saved accounts do not retain a callback listener.

Browser and protected manual input MUST share atomic, one-use admission for the exact provider/session/state. Invalid, stale and duplicate input cannot replace an accepted callback. CPA reception permits retry after failed callback-file publication. HiRoute does not resend an admitted manual code after a failed or uncertain submission: observe status, or cancel and start a new sign-in. The browser acknowledges reception without claiming token exchange or sign-in success. Release listeners after reception, cancellation, expiry or failure, before slow token exchange. Do not redirect or reflect callback secrets in browser responses.

A port conflict MUST close partial listeners and retain the same pending session for manual completion. Desktop explains this mode before the user explicitly opens the browser, and accepts the full return URL through the existing protected input. Automatic reception opens the browser normally and retains manual completion as a fallback. Successful sign-in still requires explicit Check/Save; it does not change enabled-account or routing limits.

Requirement: Borrowed subscription refresh cannot enlarge saved membership

Only approved, saved, enabled NativeBorrowed sources may maintain their association to trusted CPA current in the same native context. Current access-token rotation and capability revocation stay inside CPA; source qualification, Gateway Attempts, and audit do not persist/compare that token generation. Stable source, selected models, Bindings, and route references remain intact.

The latest successful pinned account directory constrains original selected models; removed eligibility retains configuration as unavailable, and newly visible models are not auto-imported. One independently sealed Codex connection may preserve native models only when the user explicitly enabled it and the saved account, native directory, and CPA pinned snapshot all prove those models. This connection-only grant cannot enlarge general source selection, plan candidates, or another Agent's access. Absent proof prevents a global provider switch.

Qualification converges through protected checks, management reads, or maintenance; Attempt issuance does not discover online directories. Unobserved provider permission changes may reach upstream and use its final rejection, without enlarging selection. Unsafe credential reads, deletion, or stable account mismatch fail closed before replacement credentials or capability issuance. Revision conflicts and transient CPA/network failures retain configuration and report retry/repair. Deleted/disabled sources cannot be revived by late work, and sent requests cannot be recalled.

Requirement: Borrowed Codex auth follows its supported file semantics

An explicit HIROUTE_CODEX_AUTH_SOURCE selects that file; otherwise borrowing uses the selected CODEX_HOME file store. keyring, auto, ephemeral, and unknown store modes cannot fall back to stale auth.json; recommend independent sign-in. This path does not read Codex Keychain, decrypt native stores, or reconstruct arbitrary launch layers. Claude retains its supported file/macOS Keychain adapter; Keychain interaction is limited to explicit checking rather than passive discovery.

Explicit auth_mode=chatgpt takes precedence even when an API Key is present. Absent/null mode implies ChatGPT only when API Key/PAT/Bedrock login fields do not select another mode. Current access token must be nonempty, bounded, and usable as a request credential. id_token and last_refresh are not required for borrowing. Account ID comes first from tokens.account_id, then the supported account claim in ID/access tokens. Missing account identity requests re-login rather than using email, user ID, or token hash. Parsing claims establishes local association, not signature validation or upstream authentication. CPA-owned refresh credentials retain their own required fields.

Requirement: CPA isolation and revocation remain available during slow I/O

The packaged CPA locator binds the adjacent executable to its final signed-byte identity; missing, changed, or nonexecutable CPA affects subscriptions only and cannot silently use a user's different CPA. Native API and control remain usable. Provider failures are isolated from each other.

Credential/profile/file/Keychain/management-HTTP/process I/O MUST NOT synchronously block Gateway async workers or hold their blocking locks. Bounded executors limit running and queued work using the original absolute deadline/cancellation. Cancel cannot release a slot while its blocking work still runs. Admission revocation is separate from slow lifecycle serialization: disable/account/source switch revokes old epochs before stopping old writers, and failed cleanup remains retryable. Prepared results recheck source/account/token version/admission epoch; stale success or failure cannot overwrite newer state. Final send rechecks capability epoch.

CPA-managed authentication may refresh and retry once for the same account/model on 401 before response commitment. NativeBorrowed supplies no refresh token. HiRoute does not implement or stack another OAuth refresh chain.

Requirement: Proxy configuration is scoped and reports actual application

Subscription proxy modes are inherited startup environment, direct, and manual HTTP(S). Desired configuration persists separately from the applied process state. Manual mode consistently overrides both cases of inherited proxy variables; direct clears them; inherited mode preserves the host's startup contract. Manual URLs cannot contain credentials and inherited secrets cannot enter status/errors.

Only managed subscriptions and Claude's subscription remote checks use this policy; Native API, Worker, and system settings do not. Local management stays direct and proxy failure cannot fall back to direct internet. Editing during scan retains results, cancel returns to them, and explicit save/restart creates a fresh scan. A still-running old process or invalid configuration cannot be reported as applied.

Requirement: Model management joins facts without creating a second authority

Search/filter/detail and operation actions MUST use shared exact source/Binding, capability, reference-score, price, and subscription facts. Active route references include regular and primary candidates in smart saving/custom branches, match the Binding rather than name, and open the exact plan; disabled routes are not active use.

Saved subscriptions distinguish updating authorization, sign-in needed, connector/ network failure, and model exclusion from the latest pinned directory. They never ask for an API Key to repair OAuth. Explicit recheck uses the complete existing check/save/activation chain without reselecting membership or claiming inference.

Desktop uses ListComputeV3/GetComputeV3 and hiroute.compute-management-snapshot/v3; mode joins by source ID/revision from the committed selection even offline. Unknown selection does not guess a mode. Published ListCompute/GetCompute keep their strict V2 response fields from the same projection; released CLI stays on that allowlist. The extension changes no request envelope or persisted source and requires no implicit mode migration.

Requirement: Model feed status and origin are visible and controllable

Model management in macOS Desktop, the Windows shell, web console and public CLI SHALL expose the same daemon-owned complete model metadata and feed state. Directory search/listing, model detail, editable prefill, capability views and price views SHALL use one selected effective snapshot per operation. Every maintained descriptive record SHALL be discoverable even without an executable binding or canonical match. Bounded pagination or another bounded complete-query path SHALL preserve snapshot identity and provenance; it SHALL not silently omit records to fit a frame.

Status SHALL disclose retrieval enabled/disabled, held publication sequence/issue time, selected official data revision and origin, highest accepted publication/data revisions, last retrieval attempt/result and last success, validation/recovery state, and counts that distinguish usable metadata from execution-restricted records. An updated existing model SHALL not be counted as ignored merely because it is embedded. Record/field views SHALL distinguish website/cache/bundled defaults and explicit user overrides, retain actual reviewed-as-of information, and name unknown or unsupported components. Publication, retrieval and model-provider release times SHALL not be presented as interchangeable.

Availability SHALL distinguish executable, needs a saved connection/qualification, unsupported or needs a known newer client, withdrawn/unavailable, and use by existing published Plans. Being used by a frozen Plan SHALL not imply availability for a new publication. Descriptive-only records can be inspected but SHALL not be offered as executable. Featured entries SHALL resolve against the selected catalog and retain their editorial role and actual origin.

Every host and the CLI SHALL support explicit refresh and the same persistent automatic-retrieval setting. Disabling automatic retrieval SHALL stop scheduled requests and invalidate an earlier in-flight result without discarding the selected usable cache; a later explicit refresh remains available. Removing the copy SHALL require a preview of changed defaults, saved sources affected in new use, embedded fallback and unchanged published revisions; only confirmation of that current preview may remove it. Retained acceptance floors and user overrides SHALL survive removal. The preview SHALL account for existing models whose defaults came from the website as well as feed-only models.

Public CLI help and schema discovery SHALL describe the shared views and controls. A first display of website-derived data SHALL disclose hiroute.ai retrieval and how to turn it off. Ordinary reads SHALL neither fetch metadata, call a model, perform connection checks nor mutate configuration. Concurrent reads SHALL report internally consistent snapshot identity or an explicit refresh/conflict outcome.

Scenario: Inspect a descriptive-only model

  • WHEN the catalog contains a provider-scoped model with no executable binding
  • THEN directory search and detail expose its metadata and limitations in the shared UI and CLI without enabling an executable selection.

Scenario: Inspect mixed default and manual origins

  • WHEN an existing model uses website capabilities, a manual context limit and an unknown cache price
  • THEN the view distinguishes these origins and states rather than applying one misleading origin badge to all fields.

Scenario: Same state across hosts

  • WHEN the same service is opened through Desktop, Windows shell, web console or the public CLI
  • THEN each reports the same selected snapshot and retrieval setting; hosts do not fetch their own catalogs.

Scenario: Remove a copy that updated an embedded model

  • WHEN a held snapshot changed defaults for an embedded model used by a saved connection
  • THEN removal preview identifies the fallback/default changes and any new-use restrictions, preserves manual values and explains that published revisions remain frozen.

Scenario: Turn retrieval off

  • WHEN a user disables retrieval while a valid snapshot is held
  • THEN all clients of that service show retrieval off, the copy remains usable, and a late in-flight fetch cannot reverse the choice.

Scenario: Complete bounded enumeration

  • WHEN a caller enumerates a maintained catalog larger than one response page while a new snapshot activates
  • THEN it can complete a coherent snapshot or receives an explicit refresh requirement; it is never told that a silently truncated or mixed-generation list is complete.

Scenario: Refresh updates only selected source facts

  • WHEN a publication updates one provider while other records retain older evidence dates
  • THEN the views show the new publication/retrieval time separately from each record's actual reviewed-as-of date.

Source SPEC comments:

Implementation references