Foundation catalog drift DD-1/DD-2: one Agents SDK disposition and the current gateway builds

catalogs/foundation/decisions.json

@@ -2842,12 +2842,12 @@
       ],
       "evidence_scope": "Reuses the existing native Codex CLI/SDK inventory and acceptance evidence; no new execution was run to freeze this taxonomy layer.",
       "limitations": [
-        "Claude Agent SDK, OpenHands SDK, Temporal and LangGraph remain unqualified candidates here; their version metadata is source review only, not fresh execution."
+        "Claude Agent SDK, OpenHands SDK, Temporal and LangGraph remain unqualified candidates here; their version metadata is source review only, not fresh execution. The 2026-10-03 source revisit also records OpenAI Agents SDK as unqualified; its experimental codex_tool and SDK requests through OmniRoute have no measured qualification."
       ],
       "source_paths": [
         "docs/foundation-closure-20260921.md"
       ],
-      "next_gap": "The SDK table in docs/foundation-closure-20260921.md lines 78-84 has not been revisited by any lane since 2026-09-21; qualify a concrete application need before promoting Claude Agent SDK, OpenHands SDK, Temporal or LangGraph.",
+      "next_gap": "The 2026-09-21 SDK table in docs/foundation-closure-20260921.md has its header at line 80 and candidate rows at lines 82-86, with no OpenAI Agents SDK row. The 2026-10-03 source revisit corrects the table citation and records OpenAI Agents SDK as unqualified in the current catalogs; it supplies no new execution. Qualify a concrete application need before promoting Claude Agent SDK, OpenHands SDK, Temporal, LangGraph or OpenAI Agents SDK; the latter's codex_tool adoption hold and newly preregistered comparison path are defined in docs/decisions/2026-10-03-foundation-catalog-drift.md.",
       "lifecycle": {
         "source_path": "blueprints/token-native-focus/saturation-audit.json",
         "scope": "Only the listed canonical stage and its original host/operation scope apply. No new execution was run for this taxonomy freeze.",

catalogs/foundation/manifest.json

@@ -20,7 +20,7 @@
       "id": "native-clients",
       "title": "Native clients",
       "purpose": "Run ordinary coding, research and review through locally signed-in native clients.",
-      "selection": "Codex and Claude keep native accounts, model selection, caching and compaction. Model routing and providers are not a separate layer here, and the catalogs/us-equities/models.json ledger remains a note on this native-clients selection. The user decided on 2026-09-27 to adopt a loopback OmniRoute gateway (component omniroute) for pooled GPT-6 access by gateway-routed Codex lanes such as the landscape sweep. It is a source build of release/v3.8.51 with an upstream fix (#14904) and an upstream feature (#13788); see docs/decisions/2026-09-27-omniroute-account-pool.md. Native Codex stays the max-quality default. The agent-sdks layer default stays pending the preregistered three-arm workers comparison (docs/grand-catalog-handbook.md, agent-sdks row), which remains its only acceptance gate. The terminal surface of these clients (tab titles, bell, colour depth, needed-only alerts) is settled by docs/decisions/2026-09-28-terminal-experience.md: AI-client (Claude, Codex) Windows Terminal profiles must not suppress application titles (static shell profiles may keep fixed titles), BEL is the only bell or notification signal Windows Terminal 1.24 and 1.25 act on (OSC 9;4 only sets progress state), and alerts are limited to pending decisions.",
+      "selection": "Codex and Claude keep native accounts, model selection, caching and compaction. Model routing and providers are not a separate layer here, and the catalogs/us-equities/models.json ledger remains a note on this native-clients selection. The user decided on 2026-09-27 to adopt a loopback OmniRoute gateway (component omniroute) for pooled GPT-6 access by gateway-routed Codex lanes such as the landscape sweep. Dated workstation observations reconciled on 2026-10-03 identify 20128 as omniroute-3.8.51-2f42a9ac-pr13788-affinity2-pr15167 (BUILD_SHA cf6748d04) and 20129 as omniroute-3.8.51-2f42a9ac-pr13788 (BUILD_SHA 87c4c488d). Both use base 2f42a9ac19d1a247ec9ce5473b790843724b3061, whose tree matches v3.8.51. Both carry upstream #13788 (head 6c7990058c4ce9677de79452c8cefb10b4bf1b3d); 20128 also carries the local affinity patch and upstream #15167 cherry-picked from f5d8e150b79e0901fa18241c7f29bff889b87c14 into build cf6748d04, as retained in evidence/artifacts/omniroute-sol-max-20260930/checks/qualification-pipeline-times.txt and evidence/artifacts/omniroute-sol-max-20260930/checks/omniroute.service.after-switch.txt. The PR's later force-pushed current head 0585aba5589d5a1f49243a13a8db249558e7c9e3 is current upstream state only; its reasoningSuffix.ts is byte-identical to the carried head. The clean omniroute@3.8.51 package caps gpt-6.1-sol at xhigh; the carried PR #15167 keeps requested max, and the Codex executor maps ultra to wire max. Sources: docs/decisions/2026-09-30-omniroute-rebuild.md, PR #637's decision at https://github.com/seathatflowsinourveins/native-agent-stack/blob/269f9516883b/docs/decisions/2026-10-03-omniroute-3851-pin.md, and pinned upstream effort sources linked in docs/decisions/2026-10-03-foundation-catalog-drift.md, including https://github.com/diegosouzapw/OmniRoute/blob/f5d8e150b79e0901fa18241c7f29bff889b87c14/open-sse/executors/codex/reasoningSuffix.ts#L11-L33 and the current upstream state at https://github.com/diegosouzapw/OmniRoute/blob/0585aba5589d5a1f49243a13a8db249558e7c9e3/open-sse/executors/codex/reasoningSuffix.ts#L11-L33. These descriptions are dated observations, not new host acceptance. Native Codex stays the max-quality default. The agent-sdks layer default stays pending the preregistered three-arm workers comparison (docs/grand-catalog-handbook.md, agent-sdks row), which remains its only acceptance gate. The terminal surface of these clients (tab titles, bell, colour depth, needed-only alerts) is settled by docs/decisions/2026-09-28-terminal-experience.md: AI-client (Claude, Codex) Windows Terminal profiles must not suppress application titles (static shell profiles may keep fixed titles), BEL is the only bell or notification signal Windows Terminal 1.24 and 1.25 act on (OSC 9;4 only sets progress state), and alerts are limited to pending decisions.",
       "activation": "Use the intended native client and inspect its loaded project tools; new process defaults require a fresh session.",
       "lifecycle_scope": "Dated native source retrieval, process environment and scoped session continuation passed; host-wide restart recovery is separate.",
       "next_gap": "Bind fresh-session activation and resumed effects to the next target host without repeating accepted model trials."
@@ -164,10 +164,10 @@
       "id": "agent-sdks",
       "title": "Agent SDKs and runtime workers",
       "purpose": "Select the programmatic SDK or runtime that gives a worker structured events, custom tools or session control beyond the native CLI.",
-      "selection": "Retain native Codex CLI/SDK for session work. Use an explicitly configured OmniRoute SDK lane for application tools/events/handoffs; Pi remains an optional trial. The source-only 2026-10-02 two-host review is in docs/decisions/2026-10-02-two-host-north-star-architecture.md; it does not promote a new runtime.",
+      "selection": "Retain native Codex CLI/SDK for session work. Use an explicitly configured OmniRoute SDK lane for application tools/events/handoffs; Pi remains an optional trial. OpenAI Agents SDK v0.23.1 is an unqualified alternative: OPENAI_BASE_URL supports an OpenAI-compatible endpoint. Experimental codex_tool forwards an explicit codex_options.api_key, otherwise resolves CODEX_API_KEY or OPENAI_API_KEY from options.env, then the process environment, then the set_default_openai_key value, and exports the resolved key as CODEX_API_KEY. The native Codex login applies only when no key resolves, or when the caller supplies a Codex instance without api_key and preserves its native authentication environment/configuration. The managed Agents API separately requires an API key and bills at API rates. SDK requests through OmniRoute have not been measured; the 2026-10-02 Codex-practice audit item AD-5, hold on Agents SDK codex_tool adoption, remains unresolved. The source-only 2026-10-02 two-host review is in docs/decisions/2026-10-02-two-host-north-star-architecture.md; it does not promote a new runtime. The pinned SDK sources and 2026-10-03 catalog correction are in docs/decisions/2026-10-03-foundation-catalog-drift.md, including https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/codex_tool.py#L584-L628, https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/codex_tool.py#L631-L639, https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/thread.py#L108-L111 and https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/exec.py#L226-L240.",
       "activation": "Use the native CLI for ordinary work; reach for an official SDK only when an application needs structured events, custom tools or programmatic session control.",
       "lifecycle_scope": "Retain each runtime's recorded host, version and task scope. Native Codex receipts, separate SDK protocol checks and the unmerged Pi trial are distinct; the 2026-10-02 source review supplies no new model, runtime or workstation lifecycle acceptance.",
-      "next_gap": "The 2026-10-02 source review distinguishes native Ultra from API reasoning enums and Pi 0.99.1 trial evidence from upstream 1.0.0. Qualify the actual application's endpoint, locked SDK, extensions, cancellation and resume before broader promotion; workstation activation remains host-specific."
+      "next_gap": "The 2026-10-02 source review distinguishes native Ultra from API reasoning enums and Pi 0.99.1 trial evidence from upstream 1.0.0. Qualify the actual application's endpoint, locked SDK, extensions, cancellation and resume before broader promotion; workstation activation remains host-specific. The 2026-10-02 Codex-practice audit item AD-5, hold on Agents SDK codex_tool adoption, remains unresolved: adoption first needs a demonstrated orchestration requirement and graduation from experimental, then an accepted newly preregistered comparison of codex_tool, the Codex thread API and codex exec --json plus resume, per catalogs/sota-convergence/manifest-20260926.json:451, with complete usage, events, external effects, cancellation and recovery. This candidate comparison stays separate from the sealed three-arm workers comparison, whose layer-default verdict changes only through a re-preregistered rerun (docs/grand-catalog-handbook.md, agent-sdks row). The frozen agent-sdks and native-clients verdict rows (wave 20260922) name no OpenAI Agents SDK overturn path, so a promotion path first needs those rows re-recorded under a new run id with a sealed cross-family review (open item in docs/decisions/2026-10-03-foundation-catalog-drift.md)."
     },
     {
       "id": "mcp-surfaces",

catalogs/landscape/foundation.json

@@ -46,12 +46,19 @@
         {
           "name": "OpenAI Agents SDK",
           "repository": "https://github.com/openai/openai-agents-python",
-          "disposition": "conditional",
-          "rationale": "Fits an application-owned API agent; it is not a substitute for the established subscription-backed native Codex account and behavior.",
+          "disposition": "unqualified",
+          "rationale": "Source review corrected on 2026-10-03. The managed Agents API requires OPENAI_API_KEY and bills at API rates. The open-source SDK supports OpenAI-compatible endpoints through OPENAI_BASE_URL. Its experimental codex_tool forwards an explicit codex_options.api_key, otherwise resolves CODEX_API_KEY or OPENAI_API_KEY from options.env, then the process environment, then the set_default_openai_key value, and exports the resolved key as CODEX_API_KEY to the local Codex CLI. The native Codex login applies only when no key resolves, or when the caller supplies a Codex instance without api_key and preserves its native authentication environment/configuration. SDK requests through OmniRoute have not been measured. The 2026-10-02 Codex-practice audit item AD-5, hold on Agents SDK codex_tool adoption, remains unresolved: no demonstrated orchestration requirement or accepted newly preregistered comparison of codex_tool, the Codex thread API and codex exec --json plus resume, and codex_tool is still experimental.",
           "evidence_kind": "source_review",
           "evidence_refs": [
-            "docs/foundation-closure-20260921.md",
-            "catalogs/us-equities/agents-operations.json"
+            "docs/decisions/2026-10-03-foundation-catalog-drift.md",
+            "https://developers.openai.com/api/docs/guides/agents-api/overview",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/codex_tool.py#L584-L628",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/codex_tool.py#L631-L639",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/thread.py#L108-L111",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/exec.py#L226-L240",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/docs/config.md#L84-L88",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/docs/tools.md#L890-L894",
+            "catalogs/sota-convergence/manifest-20260926.json"
           ]
         },
         {
@@ -4558,7 +4565,7 @@
       "requirement": "Give a worker structured events, custom tools or session control through a supported SDK when the native CLI's bounded review is not enough.",
       "current_choice": "Codex SDK/CLI (component codex) for supported start/run/resume; Claude Agent SDK, OpenHands SDK, Temporal and LangGraph remain source-reviewed, unselected alternatives.",
       "decision": "retain",
-      "rationale": "Codex SDK integration is accepted only within its recorded tests and needs no new runtime; the SDK table in docs/foundation-closure-20260921.md (lines 78-84) reviewed the alternatives on 2026-09-21 and none has since been revisited by any lane.",
+      "rationale": "Codex SDK integration is accepted only within its recorded tests and needs no new runtime. The 2026-09-21 SDK table in docs/foundation-closure-20260921.md has its header at line 80 and five candidate rows at lines 82-86, with no OpenAI Agents SDK row. The 2026-10-03 source revisit adds that SDK as unqualified without qualifying its experimental codex_tool or its requests through OmniRoute; see docs/decisions/2026-10-03-foundation-catalog-drift.md.",
       "evidence_refs": [
         "docs/foundation-closure-20260921.md",
         "catalogs/foundation/decisions.json"
@@ -4580,6 +4587,25 @@
             "docs/foundation-closure-20260921.md"
           ]
         },
+        {
+          "name": "OpenAI Agents SDK",
+          "repository": "https://github.com/openai/openai-agents-python",
+          "disposition": "unqualified",
+          "rationale": "Source review at v0.23.1 supports an OpenAI-compatible SDK endpoint. Experimental codex_tool forwards an explicit codex_options.api_key, otherwise resolves CODEX_API_KEY or OPENAI_API_KEY from options.env, then the process environment, then the set_default_openai_key value, and exports the resolved key as CODEX_API_KEY. The native Codex login applies only when no key resolves, or when the caller supplies a Codex instance without api_key and preserves its native authentication environment/configuration. The managed Agents API separately requires an API key and bills at API rates. SDK requests through OmniRoute have not been measured. The 2026-10-02 Codex-practice audit item AD-5, hold on Agents SDK codex_tool adoption, stays on hold until a needed orchestration capability, graduation from experimental and an accepted newly preregistered comparison of codex_tool, the Codex thread API and codex exec --json plus resume justify adoption, with complete usage and effect/recovery checks. This candidate comparison is separate from the sealed three-arm workers comparison. The completed two-family fit refutation is retained at evidence/artifacts/landscape-sweep-20260926/returns.json#/votes/agent-sdks/6/fit; correcting the Claude vote's managed-API premise does not qualify the candidate.",
+          "evidence_kind": "source_review",
+          "evidence_refs": [
+            "docs/decisions/2026-10-03-foundation-catalog-drift.md",
+            "evidence/artifacts/landscape-sweep-20260926/returns.json",
+            "catalogs/sota-convergence/manifest-20260926.json",
+            "https://developers.openai.com/api/docs/guides/agents-api/overview",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/codex_tool.py#L584-L628",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/codex_tool.py#L631-L639",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/thread.py#L108-L111",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/src/agents/extensions/experimental/codex/exec.py#L226-L240",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/docs/config.md#L84-L88",
+            "https://github.com/openai/openai-agents-python/blob/v0.23.1/docs/tools.md#L890-L894"
+          ]
+        },
         {
           "name": "Claude Agent SDK",
           "repository": "https://github.com/anthropics/claude-agent-sdk-python",

catalogs/landscape/research-state.json

@@ -387,7 +387,7 @@
       "catalog": "foundation",
       "layer_id": "agent-sdks",
       "status": "on_requirement_change",
-      "next_action": "Qualify a concrete application need (durable multi-host state, explicit graph checkpoints, a remote/container worker or programmatic Claude session control) before promoting a deferred SDK alternative; the SDK table in docs/foundation-closure-20260921.md (lines 78-84) has not been revisited by any lane since 2026-09-21.",
+      "next_action": "Qualify a concrete application need (durable multi-host state, explicit graph checkpoints, a remote/container worker or programmatic session control) before promoting a deferred SDK alternative. The 2026-09-21 SDK table in docs/foundation-closure-20260921.md has its header at line 80 and candidate rows at lines 82-86, with no OpenAI Agents SDK row. The 2026-10-03 source revisit corrects the citation and adds OpenAI Agents SDK as unqualified in the current catalogs; it supplies no new execution and preserves the codex_tool adoption hold and newly preregistered comparison path in docs/decisions/2026-10-03-foundation-catalog-drift.md.",
       "decision_ref": "catalogs/landscape/foundation.json",
       "evidence_refs": [
         "docs/foundation-closure-20260921.md",