feat(profiles): promote codex's full tier to GPT-6 Astra

src/model_profile_trial.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Guarded, read-only Sol/Terra/Luna worker-profile plumbing trial.
+"""Guarded, read-only Astra/Terra/Luna worker-profile plumbing trial.
 
 The trial is instrumentation, not a benchmark: one frozen packet, one shared
 Codex capacity snapshot, randomized launch order, and three exact profiles. It
@@ -55,7 +55,7 @@ def _manifest_profile_ids(manifest: dict[str, Any]) -> tuple[str, ...]:
     for supported in (EXPECTED_PROFILE_IDS, LEGACY_PROFILE_IDS):
         if actual == set(supported) and len(requests) == len(supported):
             return supported
-    raise ValueError("trial requires an exact supported three-profile set")
+    raise ValueError("trial requires an exact supported profile set")
 
 
 DEFAULT_STATE_PATH = Path(
@@ -370,6 +370,7 @@ def validate_trial_manifest(manifest: dict[str, Any]) -> None:
         ):
             raise ValueError("trial source manifest is incomplete")
     profile_ids = _manifest_profile_ids(manifest)
+    expected_count = len(profile_ids)
     identity = {
         "created_at": manifest.get("created_at"),
         "packet_hash": manifest.get("packet_hash"),
@@ -385,9 +386,9 @@ def validate_trial_manifest(manifest: dict[str, Any]) -> None:
 
     requests = manifest.get("requests") or []
     if {item.get("profile_id") for item in requests} != set(profile_ids):
-        raise ValueError("trial requires an exact supported three-profile set")
-    if len({item.get("run_id") for item in requests}) != 3:
-        raise ValueError("trial requires three distinct run identities")
+        raise ValueError("trial requires an exact supported profile set")
+    if len({item.get("run_id") for item in requests}) != expected_count:
+        raise ValueError("trial requires one distinct run identity per profile")
     if manifest.get("capacity_snapshot", {}).get("snapshot_count") != 1:
         raise ValueError("trial requires one shared-pool snapshot")
     launch_order = manifest.get("launch_order") or []
@@ -396,8 +397,10 @@ def validate_trial_manifest(manifest: dict[str, Any]) -> None:
     by_ordinal = sorted(requests, key=lambda item: int(item.get("launch_ordinal") or 0))
     if [item.get("profile_id") for item in by_ordinal] != launch_order:
         raise ValueError("trial request ordinals do not match launch order")
-    if [int(item.get("launch_ordinal") or 0) for item in by_ordinal] != [1, 2, 3]:
-        raise ValueError("trial launch ordinals must be exactly 1,2,3")
+    if [int(item.get("launch_ordinal") or 0) for item in by_ordinal] != list(
+        range(1, expected_count + 1)
+    ):
+        raise ValueError("trial launch ordinals must cover the selected profile set")
     for request in requests:
         profile = execution_profiles.get_profile(request["profile_id"])
         expected_run_id = f"{manifest['trial_id']}:{request['profile_id']}"
@@ -635,8 +638,8 @@ def finalize_trial(
             for request in manifest["requests"]
         ]
     )
-    if debit != {"codex-subscription": 3.0}:
-        raise AssertionError("trial did not debit one shared pool exactly three times")
+    if debit != {"codex-subscription": float(len(manifest["requests"]))}:
+        raise AssertionError("trial did not debit one shared pool once per profile")
     state = {
         "schema": STATE_SCHEMA,
         "version": SCHEMA_VERSION,

tests/rail_exercises/local-model-profile-trial/arm-a/contract.json

@@ -19,11 +19,11 @@
     "round": "r15",
     "source": "rounds/r15/proposers/P3-cursor/orch/exercises/local-model-profile-trial.json"
   },
-  "pass_check": "python3 -c \"import json,sys; m=json.load(open('$CONTRACT_DIR/fixtures/local-model-profile-trial/manifest.json')); ids={r['profile_id'] for r in m['requests']}; ok=(m.get('promotion_allowed') is False and m.get('learning_enabled') is False and ids=={'codex-5.6-sol-high','codex-5.6-terra-high','codex-5.6-luna-high'} and m.get('stop_conditions',{}).get('triggered') is False); sys.exit(0 if ok else 1)\"",
+  "pass_check": "python3 -c \"import json,sys; m=json.load(open('$CONTRACT_DIR/fixtures/local-model-profile-trial/manifest.json')); ids={r['profile_id'] for r in m['requests']}; ok=(m.get('promotion_allowed') is False and m.get('learning_enabled') is False and ids=={'codex-6-astra-high','codex-5.6-terra-high','codex-5.6-luna-high'} and m.get('stop_conditions',{}).get('triggered') is False); sys.exit(0 if ok else 1)\"",
   "run": [
     "cd $REPO_ROOT/orch && python3 src/model_profile_trial.py prepare --orchestrator-root $CONTRACT_DIR/fixtures/local-model-profile-trial/orch --workflows-root $CONTRACT_DIR/fixtures/local-model-profile-trial/workflows --output $CONTRACT_DIR/fixtures/local-model-profile-trial/manifest.json --capacity-state ok > $CONTRACT_DIR/fixtures/local-model-profile-trial/prepare-out.json",
     "cd $REPO_ROOT/orch && python3 -c \"import json,sys; sys.path.insert(0,'src'); from model_profile_trial import validate_trial_manifest; m=json.load(open('$CONTRACT_DIR/fixtures/local-model-profile-trial/manifest.json')); validate_trial_manifest(m); assert m['schema']=='orchestrator.model-profile-trial'; assert m['lifecycle']=='shadow'; assert len(m['requests'])==3; print('validate_ok')\""
   ],
   "setup": [],
-  "stated_task": "Prepare and validate a shadow Sol/Terra/Luna trial manifest from minimal fixture source trees without executing workers or recording feedback."
+  "stated_task": "Prepare and validate a shadow Astra/Terra/Luna trial manifest from minimal fixture source trees without executing workers or recording feedback."
 }

tests/rail_exercises/local-model-profile-trial/arm-a/fixtures/local-model-profile-trial/manifest.json

@@ -95,6 +95,12 @@
         ],
         "state": "ok"
       },
+      "codex-6-astra-high": {
+        "capacity_pool_ids": [
+          "codex-subscription"
+        ],
+        "state": "ok"
+      },
       "cursor-composer-2.5": {
         "capacity_pool_ids": [
           "cursor-subscription"
@@ -142,7 +148,7 @@
   "launch_order": [
     "codex-5.6-terra-high",
     "codex-5.6-luna-high",
-    "codex-5.6-sol-high"
+    "codex-6-astra-high"
   ],
   "learning_enabled": false,
   "lifecycle": "shadow",
@@ -163,7 +169,7 @@
       "provider": "openai",
       "reasoning_effort": "high",
       "requested_model": "gpt-5.6-terra",
-      "run_id": "model-profile-trial:d965f949e25603e6d56293f9:codex-5.6-terra-high",
+      "run_id": "model-profile-trial:e5267a218ea4136c9682a5ff:codex-5.6-terra-high",
       "sandbox": "read-only",
       "tools": []
     },
@@ -181,7 +187,7 @@
       "provider": "openai",
       "reasoning_effort": "high",
       "requested_model": "gpt-5.6-luna",
-      "run_id": "model-profile-trial:d965f949e25603e6d56293f9:codex-5.6-luna-high",
+      "run_id": "model-profile-trial:e5267a218ea4136c9682a5ff:codex-5.6-luna-high",
       "sandbox": "read-only",
       "tools": []
     },
@@ -194,12 +200,12 @@
       "learning_enabled": false,
       "packet_hash": "sha256:343604df7f0b27aa9df5507a34fd48beffdc1627987439c7b89e72af715a7605",
       "permission_mode": "read-only",
-      "profile_id": "codex-5.6-sol-high",
+      "profile_id": "codex-6-astra-high",
       "prompt_version": "orchestrator-prompt-v1",
       "provider": "openai",
       "reasoning_effort": "high",
-      "requested_model": "gpt-5.6-sol",
-      "run_id": "model-profile-trial:d965f949e25603e6d56293f9:codex-5.6-sol-high",
+      "requested_model": "gpt-6-astra",
+      "run_id": "model-profile-trial:e5267a218ea4136c9682a5ff:codex-6-astra-high",
       "sandbox": "read-only",
       "tools": []
     }
@@ -236,6 +242,6 @@
     "reasons": [],
     "triggered": false
   },
-  "trial_id": "model-profile-trial:d965f949e25603e6d56293f9",
+  "trial_id": "model-profile-trial:e5267a218ea4136c9682a5ff",
   "version": 1
 }

tests/test_capacity_profiles.py

@@ -161,6 +161,19 @@ def test_nested_sandbox_never_widens_read_only_profile(monkeypatch):
     assert command[command.index("--sandbox") + 1] == "read-only"
 
 
+def test_full_tier_selects_astra_and_keeps_historical_sol_lookup():
+    assert adapters.resolve_model("codex", "full") == "gpt-6-astra"
+    active = execution_profiles.profiles_for_agent("codex")
+    full = [p for p in active if p["requested_model"] == adapters.resolve_model("codex", "full")]
+    assert [p["profile_id"] for p in full] == ["codex-6-astra-high"]
+    assert "codex-5.6-sol-high" not in {p["profile_id"] for p in active}
+    selected = execution_profiles.select_profile(
+        "implement", "o/r#full-tier", [p["profile_id"] for p in full], rng_seed=0
+    )
+    assert selected["selected_profile_id"] == "codex-6-astra-high"
+    assert execution_profiles.get_profile("codex-5.6-sol-high")["requested_model"] == "gpt-5.6-sol"
+
+
 def test_router_profile_envelope_replays_exact_choice():
     cap = capacity.profile_capacity_snapshot(
         {

tests/test_model_profile_trial.py

@@ -1,6 +1,10 @@
 from __future__ import annotations
 
 import json
+import os
+import shutil
+import subprocess
+from pathlib import Path
 
 import pytest
 
@@ -325,3 +329,50 @@ def test_frozen_sol_trial_remains_valid_after_astra_migration(trial_roots, monke
     )
     assert "codex-6-astra-high" in fresh["launch_order"]
     assert fresh["trial_id"] != frozen["trial_id"]
+
+
+@pytest.mark.parametrize("profile_count", [2, 4])
+def test_manifest_cardinality_follows_selected_profiles(trial_roots, monkeypatch, profile_count):
+    profiles = (
+        "codex-6-astra-high",
+        "codex-5.6-terra-high",
+        "codex-5.6-luna-high",
+        "codex-5.6-sol-high",
+    )[:profile_count]
+    monkeypatch.setattr(model_profile_trial, "EXPECTED_PROFILE_IDS", profiles)
+    manifest = model_profile_trial.build_trial_manifest(
+        *trial_roots, seed=14, now=1000, capacity_state="ok"
+    )
+    model_profile_trial.validate_trial_manifest(manifest)
+    results = trial_attempt_fixture.__wrapped__(manifest)
+    assert len(model_profile_trial._validate_results(manifest, results)) == profile_count
+    state = model_profile_trial.finalize_trial(manifest, results, now=1000)
+    assert state["shared_pool_debit"] == {"codex-subscription": float(profile_count)}
+    assert state["attempt_count"] == profile_count
+
+
+def test_committed_rail_contract_accepts_current_astra_trial(tmp_path):
+    source = Path(__file__).parent / "rail_exercises/local-model-profile-trial/arm-a"
+    contract = json.loads((source / "contract.json").read_text())
+    fixture = source / "fixtures/local-model-profile-trial/manifest.json"
+    frozen = json.loads(fixture.read_text())
+    assert {r["profile_id"] for r in frozen["requests"]} == {
+        "codex-6-astra-high",
+        "codex-5.6-terra-high",
+        "codex-5.6-luna-high",
+    }
+    model_profile_trial.validate_trial_manifest(frozen)
+    copied = tmp_path / "arm-a"
+    shutil.copytree(source, copied)
+    roots = copied / "fixtures/local-model-profile-trial"
+    fresh = model_profile_trial.build_trial_manifest(
+        roots / "orch", roots / "workflows", seed=14, now=1000, capacity_state="ok"
+    )
+    (roots / "manifest.json").write_text(json.dumps(fresh))
+    result = subprocess.run(
+        ["bash", "-eu", "-c", contract["pass_check"]],
+        env={**os.environ, "CONTRACT_DIR": str(copied)},
+        capture_output=True,
+        text=True,
+    )
+    assert result.returncode == 0, result.stderr