release: prepare 1.2.2 for Codex 0.145

Author: monarchjuno

CHANGELOG.md

@@ -2,6 +2,15 @@
 
 ## Unreleased
 
+## 1.2.2 - 2026-07-22
+
+- Validate Codex CLI 0.145.0 as the current reference while retaining 0.144.4
+  as the compatibility floor; adopt its stable MultiAgent V2 schema by leaving
+  nesting and concurrency to native Codex while preserving strict config,
+  permission, hook, MCP, and role-model projection checks.
+- Restore the Head Manager's generated `export_dataset_csv` projection by
+  registering its canonical MCP schema, role boundary, and export capability,
+  with a generated-config-to-registry regression check.
 - Stop automated GitHub Release notes at the next changelog heading so a
   release contains only its matching version section.
 

docs/codex-native-orchestration.md

@@ -81,7 +81,8 @@ conflicts with Strategy.
 
 ## Delegation Contract
 
-MultiAgent V2 exposes the generated role profiles with a depth-one boundary.
+MultiAgent V2 exposes the generated role profiles through native Codex
+delegation.
 Head Manager prefers an exact profile when its specialty is useful and passes a
 compact role-owned brief rather than the full root history. Role TOML supplies
 the direct fixed model settings; their values and fallback eligibility are owned
@@ -98,11 +99,15 @@ Use `followup_task` when a live child still owns a correction or clarification.
 Start another child for a new specialty, an unavailable session, or independent
 review. Apply the canonical fallback boundary in
 [Roles, Skills, And Workflows](roles-skills-and-workflows.md). A child cannot
-approve, execute, access a broker, or emulate Head Manager. Children never
-delegate recursively. Head Manager waits only while a live child has useful work
-and reports lifecycle events only from native tool and child-lifecycle results
-in the current run. Native wait is wait-any and may serialize with no explicit
-target list; that representation is not failure by itself.
+approve, execute, access a broker, or emulate Head Manager. Whether to delegate
+a distinct child-owned subtask further is a native Codex decision, not a
+TradingCodex workflow requirement. Every descendant remains subject to its
+selected projected role and the shared evidence, authority, and safety
+boundaries. Head Manager
+waits only while a live child has useful work and reports lifecycle events only
+from native tool and child-lifecycle results in the current run. Native wait is
+wait-any and may serialize with no explicit target list; that representation is
+not failure by itself.
 
 When a Brain applies, the assignment contains the question Head Manager derived
 from it, not the Brain body or a delegation of Brain authority. Brain content is

docs/generated-workspaces.md

@@ -534,7 +534,7 @@ That explicit overlay rule applies to external skill procedures, not to
 read-only app, connector, MCP, or data tools used as evidence. A role whose
 assignment needs external data or preserves a user-named provider first checks
 the current task's callable tool surface and uses the runtime's available
-deferred-tool discovery surface when needed. The Codex 0.144.4 reference
+deferred-tool discovery surface when needed. The Codex 0.145.0 reference
 contract first resolves at most 12 names with
 `text(ALL_TOOLS.filter(x => x.name.includes("<provider-or-keyword>")).slice(0, 12).map(x => x.name))`.
 One supported query may combine one to four literal `name.includes`
@@ -586,14 +586,22 @@ runs, hard-failure checks, blind review, and resolved outcomes.
 Generated `.codex/config.toml` enables MultiAgent V2 with visible spawn
 metadata and the `agents` tool namespace. It keeps every
 `.codex/agents/*.toml` role discoverable while setting
-`features.multi_agent_v2.enabled = true` and `agents.max_depth = 1`. The
-V1-only `agents.max_threads` key is absent because Codex rejects it when V2 is
-enabled. Codex owns concurrency and scheduling. Head Manager chooses useful
-profiles and follows up with an owner when appropriate. Role eligibility and
-fallback are defined in [Roles, Skills, And Workflows](roles-skills-and-workflows.md).
-Children cannot recursively dispatch. Root model and reasoning settings are
-inherited, and direct fixed-role TOML model settings are projected, but `doctor`
-does not duplicate native model availability or lifecycle checks.
+`features.multi_agent_v2.enabled = true`. TradingCodex does not set
+`agents.max_depth`, `agents.max_threads`, or
+`agents.max_concurrent_threads_per_session`; Codex 0.145.0 ignores
+`agents.max_depth` for V2, retains `agents.max_threads` as a compatibility
+alias, and owns nesting, concurrency, and scheduling. Head Manager chooses
+useful profiles and follows up with an owner when appropriate. Role eligibility
+and fallback are defined in [Roles, Skills, And Workflows](roles-skills-and-workflows.md).
+Whether to delegate distinct child-owned work further is a native Codex
+decision, not a TradingCodex workflow requirement. Every descendant remains
+subject to its selected projected role and TradingCodex safety and service
+gates. Root model and reasoning settings are inherited, and direct fixed-role
+TOML model settings are projected, but `doctor` does not duplicate native model
+availability or lifecycle checks. Codex 0.145.0 exposes optional per-spawn model
+and reasoning overrides by default; TradingCodex neither disables nor requires
+them. A spawn without an explicit override follows native inheritance and role
+config resolution.
 
 Workspace template modules are deployment projections. Agent and skill
 ownership comes from the Python agent registry and is projected
@@ -868,10 +876,11 @@ Generated root config sets `default_permissions = "trading-research"` and
 defines two custom profiles. It deliberately omits legacy `sandbox_mode` from
 the root and every fixed-role TOML because any loaded `sandbox_mode` overrides
 custom permission profiles in Codex. This contract requires Codex CLI 0.144.4
-or later on a locally supported platform. Version 0.144.4 is the current
+or later on a locally supported platform. Version 0.145.0 is the current
 release-validation reference for permission profiles, hooks, required MCP,
-deferred MCP calls, and the explicit V2 feature table. Older versions fail the
-Codex runtime doctor check. These remain version-sensitive
+deferred MCP calls, and the explicit V2 feature table. Versions below the
+compatibility floor fail the Codex runtime doctor check; compatible non-reference
+versions warn. These remain version-sensitive
 surfaces, so release validation includes strict config/feature inspection and
 a real native smoke.
 See the [Codex permissions reference](https://learn.chatgpt.com/docs/permissions).
@@ -1582,7 +1591,7 @@ Codex-native bootstrap verification:
   workspace. The canonical workspace must
   either be interactively trusted or receive the one-run `projects={...}` trust
   override shown in `AGENTS.md`. A full fixed-role lifecycle smoke must use
-  persisted trust for all five generated project hooks. Codex 0.144.4 does not
+  persisted trust for all five generated project hooks. Codex 0.145.0 does not
   carry the one-run `--dangerously-bypass-hook-trust` flag through the role
   config reload used by an exact V2 child, so that flag may be used only for
   root/config diagnostics and is not lifecycle acceptance. It also does not

docs/roles-skills-and-workflows.md

@@ -226,8 +226,11 @@ Start a fresh child for a new specialty, an unavailable session, or independent
 review. The Dynamic Workflow fallback boundary applies here. A generic child
 retains the same research-only scope, evidence standard, no-secret boundary,
 and no-order boundary; it cannot approve, execute, access a broker, or act as
-Head Manager. Children never delegate recursively. Wait only while a live child
-has useful work. Native
+Head Manager. Whether to delegate a distinct child-owned subtask further is a
+native Codex decision, not a workflow requirement. Each descendant remains
+subject to its selected projected role and the shared evidence, authority, and
+safety boundaries. Wait only while a live child has useful work.
+Native
 wait-any may serialize without explicit targets; verify lifecycle through the
 native tool result and child session events rather than treating that as failure.
 
@@ -260,7 +263,7 @@ same contract error repeats.
 The same bounded rule applies across role MCP use: a documented terminal
 outcome such as stored, updated, existing, reused, or prepared ends that call;
 an unchanged deterministic failure is never blindly resubmitted. Under the
-Codex 0.144.4 deferred-tool contract, unknown-provider resolution starts with
+Codex 0.145.0 deferred-tool contract, unknown-provider resolution starts with
 the canonical names-only query
 `text(ALL_TOOLS.filter(x => x.name.includes("<provider-or-keyword>")).slice(0, 12).map(x => x.name))`.
 One names-only query may combine at most four literal `name.includes`

docs/validation-and-test-plan.md

@@ -266,8 +266,9 @@ Smoke coverage should verify:
   Manager and role-owned live research for the six evidence producers without
   treating a host finance skill as a dependency
 - generated `.codex/config.toml` explicitly enables MultiAgent V2, reports it
-  enabled through `codex features list`, leaves concurrency to Codex, and omits
-  the incompatible V1 `agents.max_threads` key
+  enabled through `codex features list`, leaves nesting and concurrency to
+  Codex, and omits TradingCodex-owned `agents.max_depth`,
+  `agents.max_threads`, and `agents.max_concurrent_threads_per_session`
 - skill/projection manifests identify the finite managed inventory, declare
   runtime discovery incomplete, and resolve exact root/role skill paths
 - two generated workspaces have different workspace ids
@@ -816,7 +817,7 @@ stop without an order, approval, execution, cancellation, broker mutation, or
 secret action. Record an unavailable Codex/auth blocker rather than replacing this with a
 claim based only on the fake subprocess test.
 
-Reference acceptance uses Codex CLI 0.144.4. Run
+Reference acceptance uses Codex CLI 0.145.0. Run
 `python tests/codex_cli_contract.py --workspace <workspace>
 --require-reference --require-hook-trust` first, after opening the disposable
 workspace in interactive Codex and persistently trusting all five generated
@@ -826,7 +827,7 @@ configuration, readable sandbox settings, the expected enabled/disabled
 feature states, and trusted lifecycle hooks. Native `codex exec` smokes must
 also pass `--strict-config`, so a newly unknown or removed project key fails
 before model behavior is accepted. Do not use `--ignore-user-config` or
-`--dangerously-bypass-hook-trust` for lifecycle acceptance: in 0.144.4 the
+`--dangerously-bypass-hook-trust` for lifecycle acceptance: in 0.145.0 the
 one-run bypass is not inherited when a V2 child reloads an exact role config.
 On platforms where a temporary directory has a symlinked alias, resolve the
 workspace once with `realpath` or `Path.resolve()` and use that same physical
@@ -839,10 +840,13 @@ Run native CLI smokes that inspect observable tool calls and child lifecycle
 results from Codex. Cover the direct fast path, one exact profile, a same-owner
 `followup_task`, and a bounded evidence-role generic fallback when a profile is
 unavailable. Briefs must stay compact, generated TOML must retain the fixed
-model settings and the project-wide `trading-research` profile, and no child may
-recursively delegate. A missing independent risk or judgment profile must remain
-explicit rather than becoming a generic review. Run a sequential two-child smoke plus one artifact-to-synthesis
-workflow because lifecycle failures may appear only after the first child exits.
+model settings and the project-wide `trading-research` profile. When native
+Codex delegates a child-owned subtask further, inspect the descendant lifecycle
+and verify that its effective role, model, permission profile, MCP principal,
+evidence boundary, and no-order boundary match the selected role. A missing independent risk or judgment
+profile must remain explicit rather than becoming a generic review. Run a
+sequential two-child smoke plus one artifact-to-synthesis workflow because
+lifecycle failures may appear only after the first child exits.
 
 For a full research smoke, inspect root and child JSONL as observable behavior,
 not private chain of thought. Require a compact bounded child brief; reject broad

installation.md

@@ -49,7 +49,7 @@ initialized is fine, as is an empty directory inside an existing parent
 worktree.
 
 Use Codex CLI 0.144.4 or later on macOS, Linux, WSL, or native Windows. Version
-0.144.4 is the current TradingCodex reference for custom permission profiles,
+0.145.0 is the current TradingCodex reference for custom permission profiles,
 hooks, required MCP startup, and explicit MultiAgent V2 configuration. Verify
 with `codex --version`; `./tcx doctor --layer guidance` fails below the
 compatibility floor and warns when the installed CLI is older or newer than the
@@ -375,10 +375,10 @@ Generated root work inherits the user's active Codex model and reasoning
 settings. All nine fixed subagents use `gpt-5.6-terra` with `high` reasoning.
 Final execution is service-owned and runs no model.
 MultiAgent V2 exposes exact custom-role routing through the `agents` namespace;
-each task uses `fork_turns="none"` and a fresh role-bound child. The generated
-V2 table explicitly sets `enabled = true` and `max_depth = 1`; concurrency
-capacity remains owned by native Codex and user/host configuration rather than
-a projected V1 thread cap.
+exact custom-role dispatch uses `fork_turns="none"` so Codex applies the fresh
+role-bound config layer. The generated V2 table explicitly sets `enabled = true`
+but does not set nesting or concurrency controls; native Codex and user/host
+configuration own those choices.
 Inspect `.codex/config.toml`, `.codex/agents/`, and `./tcx doctor` after an
 update to verify the generated Codex contract. There is no model-policy
 manifest, runtime model fallback, or rollout control.

tests/test_codex_cli_compat.py

@@ -14,8 +14,10 @@
         ("0.144.0", False, False),
         ("0.144.1", False, False),
         ("0.144.2", False, False),
-        ("0.144.4", True, False),
-        ("0.145.0", True, True),
+        ("0.144.4", True, True),
+        ("0.144.5", True, True),
+        ("0.145.0", True, False),
+        ("0.145.1", True, True),
     ],
 )
 def test_codex_cli_reference_version_check(
@@ -42,10 +44,12 @@ def test_codex_cli_reference_version_check(
     assert bool(result.get("warn")) is warn
     assert f"installed={installed}" in result["detail"]
     assert "required>=0.144.4" in result["detail"]
-    assert "reference=0.144.4" in result["detail"]
+    assert "reference=0.145.0" in result["detail"]
     if installed in {"0.144.1", "0.144.2"}:
         assert "older than required" in result["detail"]
-    if installed == "0.145.0":
+    if installed in {"0.144.4", "0.144.5"}:
+        assert "older than reference" in result["detail"]
+    if installed == "0.145.1":
         assert "newer client requires harness revalidation" in result["detail"]
 
 

tests/test_codex_native_orchestration.py

@@ -418,8 +418,8 @@ def test_generated_contract_inherits_root_model_and_keeps_role_profiles_optional
     assert '[features.multi_agent_v2]' in config
     assert 'enabled = true' in config
     assert "max_concurrent_threads_per_session" not in config
-    assert "max_depth = 1" in config
-    assert 'max_threads = 6' not in config
+    assert "max_depth" not in config_values.get("agents", {})
+    assert "max_threads" not in config_values.get("agents", {})
     assert "required = true" in config
     assert '"begin_analysis_run"' in config
     assert "record_workflow_plan" not in config
@@ -432,8 +432,8 @@ def test_generated_contract_inherits_root_model_and_keeps_role_profiles_optional
     assert "You are the `head-manager` agent" not in fixed_role
     assert "# Role And Safety" in fixed_role
     assert "authenticated TradingCodex MCP" in fixed_role
-    assert "depth-1 child" in fixed_role
-    assert "never spawn, coordinate" in fixed_role
+    assert "Whether to delegate" not in fixed_role
+    assert "All work performed on your behalf remains subject" in fixed_role
     assert "$tcx-source-gate" in fixed_role
     assert "Snapshot/Dataset/Artifact IDs" in fixed_role
     assert "do not duplicate or invent provider policy here" in fixed_role
@@ -484,6 +484,13 @@ def test_generated_contract_inherits_root_model_and_keeps_role_profiles_optional
         enabled_tools = set(role_config["mcp_servers"]["tradingcodex"]["enabled_tools"])
         assert role_config["model"] == "gpt-5.6-terra"
         assert role_config["model_reasoning_effort"] == "high"
+        assert role_config.get("agents", {}).get("enabled") is not False
+        assert (
+            role_config.get("features", {})
+            .get("multi_agent_v2", {})
+            .get("enabled")
+            is not False
+        )
         assert "get_research_artifact" in enabled_tools
         assert ARTIFACT_DISCOVERY_TOOLS.isdisjoint(enabled_tools)
         assert role_config["model_instructions_file"] == FIXED_ROLE_MODEL_INSTRUCTIONS
@@ -494,6 +501,23 @@ def test_generated_contract_inherits_root_model_and_keeps_role_profiles_optional
         assert "Artifact handoff:" not in role_instructions
 
 
+def test_generated_mcp_allowlists_reference_only_registered_tools(workspace: Path) -> None:
+    config_paths = [
+        workspace / ".codex/config.toml",
+        *(workspace / ".codex/agents").glob("*.toml"),
+    ]
+
+    for config_path in config_paths:
+        config = tomllib.loads(config_path.read_text(encoding="utf-8"))
+        enabled_tools = set(
+            config["mcp_servers"]["tradingcodex"]["enabled_tools"]
+        )
+        assert enabled_tools <= TOOL_REGISTRY.keys(), (
+            f"{config_path.relative_to(workspace)} references unregistered "
+            f"TradingCodex MCP tools: {sorted(enabled_tools - TOOL_REGISTRY.keys())}"
+        )
+
+
 def test_doctor_requires_fixed_role_base_and_per_role_override(workspace: Path) -> None:
     doctor = subprocess.run(
         [str(workspace / "tcx"), "doctor", "--verbose"],