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.
Rust ai-agentsai-gatewayclaude-codecodexcoding-agentcost-optimizationdeepseekllm-gatewayllm-routerlocal-firstmodel-routingrusttask-routingtauri
{
"artifact_kind": "hiroute.model-metadata-client-discovery-runs",
"as_of": "2026-10-14",
"catalog_digest_compared": "sha256:369479a2ccfa7828a846288db3ed85ccd6e83217bd6615758418560e8dcc9ee6",
"policy": {
"candidate_outputs_materialized": true,
"metadata_records_materialized": true,
"provenance_controls_scenario_eligibility": false,
"regeneration_required_for_candidate_review": true,
"scenario_eligibility_basis": "field-completeness",
"determinate_outcome_closure_script": ".agents/skills/model-metadata-maintainer/scripts/close_model_metadata.py"
},
"materialization": {
"model_metadata_record_count": 769,
"provider_metadata_record_count": 105,
"scenario_eligible_record_counts": {
"adapter-planning": 4,
"context-limit-prefill": 765,
"cost-hint-display": 233,
"credential-setup-hint": 105,
"custom-api-endpoint-prefill": 85,
"custom-api-model-prefill": 769,
"lifecycle-warning": 769,
"model-directory": 769,
"model-discovery-setup": 102,
"output-limit-prefill": 738,
"provider-directory": 105,
"reasoning-capability-prefill": 765,
"reasoning-rendering-hint": 368,
"request-limit-prefill": 741,
"vision-capability-prefill": 767
},
"unique_upstream_model_id_count": 401
},
"runs": [
{
"profile_key": "hermes-agent",
"profile_digest": "sha256:e106f23c7eeedc28bd4e99c2f42ab793ca5ab2a013c872f9d1b9b744eb8ebcc8",
"repository_url": "https://github.com/NousResearch/hermes-agent",
"commit": "205645ee424163c7b6cfc032c331c3557797497b",
"tree": "647b37309b9df46d51e1f05c87c608770e500180",
"result_digest": "sha256:bcfd1eb755962e55b589f44966da4cfd00444e5de89c692118ba8395b369f398",
"determinism_check": "byte-identical-two-runs",
"scanned_blob_count": 43,
"counts": {
"excluded_input": 1,
"lifecycle_candidate": 0,
"locator_candidate": 79,
"metadata_shape_signal": 8,
"model_candidate": 475,
"provider_candidate": 46
},
"comparison": {
"lexically_matched_model_candidate_count": 95,
"unmatched_model_candidate_count": 380
}
},
{
"profile_key": "openclaw",
"profile_digest": "sha256:bfa95ca075570177eadd5e1d000dcd3212911338520a3e46d4ee2a9040f803ba",
"repository_url": "https://github.com/openclaw/openclaw",
"commit": "f0817f23e939447755c614e29866b0467ee0d7c3",
"tree": "16bca19048ac6f7457cb7324b6cfe39a47a00a1c",
"result_digest": "sha256:412a94a5e7a12fecae63751b583e91bdeca71d157b2f32647db16e2a72fede92",
"determinism_check": "byte-identical-two-runs",
"scanned_blob_count": 156,
"counts": {
"excluded_input": 29,
"lifecycle_candidate": 131,
"locator_candidate": 60,
"metadata_shape_signal": 17,
"model_candidate": 277,
"provider_candidate": 44
},
"comparison": {
"lexically_matched_model_candidate_count": 78,
"unmatched_model_candidate_count": 199
}
}
],
"resolved_catalog_changes": [
{
"change_key": "claude-subscription-registered-capabilities",
"client_signal": "All Claude CPA models were missing executable endpoint capability joins and used the conservative text-only runtime fallback.",
"provider_resolution": "Reviewed the official Claude model configuration, thinking, effort and Haiku 4.5 specifications on 2026-10-10. Existing exact client repository scans are unchanged; no new scan or account entitlement verification is claimed.",
"catalog_result": "Bind seven reviewed Claude model identities through eight exact subscription IDs. Keep account eligibility conditional, preserve model ceilings and source reasoning facts, and expose the supported manual-thinking toggle subset separately from adaptive effort."
},
{
"change_key": "client-metadata-field-completeness",
"client_signal": "Hermes Agent and OpenClaw expose broad provider/model fields with uneven completeness across usage scenarios.",
"provider_resolution": "Source provenance is retained for audit and conflicts but does not assign a confidence tier or gate otherwise sufficient metadata.",
"catalog_result": "Materialized unified provider and model records with generic provenance_refs, field-derived completeness, and usable_for scenarios; operational execution remains separately qualified."
},
{
"change_key": "deepseek-v4-identity-split",
"client_signal": "Both clients distinguished current and legacy DeepSeek V4 identifiers; OpenClaw also exposed divergent vision and price assertions.",
"provider_resolution": "Provider documentation identifies DeepSeek V4.1 Flash and V4 Pro 0813 separately and marks legacy Flash identifiers as accepted retired aliases.",
"catalog_result": "Two model identities, binding-level alias lifecycle/replacement, and two product-scoped time-window pricing facts."
},
{
"change_key": "determinate-metadata-closure",
"client_signal": "Provider-scoped records kept unknown or partial capability, limit, lifecycle, modality, rendering, and endpoint fields even where official family documentation and the user decision allow a determinate conservative outcome.",
"provider_resolution": "Every indeterminate field is closed through a reviewed rule table: searched facts stay observed, conservative family/product inferences are recorded per field as basis=inferred with their rule_key and evidence, and explicit unsupported or not-applicable outcomes are used where the field does not apply.",
"catalog_result": "761 model records and 103 provider records are fully determinate; inference rules, field provenance, execution_fit outcomes, and the extended vocabulary are embedded in the catalog and re-verified by closure fixed-point and field validators."
},
{
"change_key": "gpt6-subscription-context-and-luna",
"client_signal": "A known gpt-6-luna subscription inventory ID fell back to the generic 131072-token non-thinking model; GPT-6.1 Sol incorrectly used the native client default as its model maximum.",
"provider_resolution": "Reviewed exact GPT-6 Astra/Sol/Luna and GPT-6.1 Sol official model pages on 2026-10-08: 1050000 context and 128000 output. Preserve client policy assertions and label the public context ceiling as an inference for the subscription; current CPA inventory remains the account gate.",
"catalog_result": "Register Luna in the existing Codex Responses offer, preserve none..max native reasoning, and keep the model ceiling separate from the unchanged 272000 default plan window. No subscription API price or full-context account acceptance is claimed."
},
{
"change_key": "gpt61-codex-subscription-projection",
"client_signal": "The authenticated Codex 0.159.0 directory advertises GPT-6.1 Sol and reasoning profiles; the previous current source only described the public API product.",
"provider_resolution": "Keep the native directory default/max-window values as source assertions and project the documented public model ceiling through an explicit subscription inference. Preserve documented low..max wire efforts, the separately inferred output bound and absent subscription cost hints. Existing pinned client scans are unchanged; no new repository scan is claimed.",
"catalog_result": "The existing Codex Responses offer gains the reviewed GPT-6.1 Sol definition and reasoning profiles. Account eligibility remains gated by current observed inventory."
},
{
"change_key": "tencent-hy-callable-ids",
"client_signal": "Both clients listed hy3, hy3-preview, and hy4-preview; Hermes also listed the same TokenHub and Token Plan product split.",
"provider_resolution": "Tencent product pages establish hy3, hy3-202608, hy3-preview, hy4-preview, 1024k Hy4 context, and Hy3 preview routing to Hy3 in the plan.",
"catalog_result": "Documented upstream IDs and six conditional product bindings; account availability and protocol qualification remain runtime-required."
}
]
}