Reviews default to codex on a kernel-pinned model

.github/workflows/advisory-review-agent.yml

@@ -62,18 +62,19 @@ on:
         required: false
         default: false
       backend:
-        description: Review backend (claude | hermes | codex).
+        description: Review backend (codex | claude | hermes).
         type: string
         required: false
-        default: claude
+        default: codex
       model:
         description: >-
-          Model for the session. Empty uses the OUTERLOOP_CLAUDE_MODEL
-          Actions variable for claude (no built-in default; the backend's
-          own default otherwise).
+          Model for the session. Empty uses the kernel's pinned model for
+          codex (harnesses.toml, moved with releases), the
+          OUTERLOOP_CLAUDE_MODEL Actions variable for claude (no built-in
+          default), and hermes's own default.
           For hermes pass an OpenRouter id — except with
           hermes_provider: openai, which REQUIRES the provider-native id
-          (gpt-5.6-terra, no openai/ prefix).
+          (gpt-x.y, no openai/ prefix).
         type: string
         required: false
         default: ''
@@ -82,7 +83,7 @@ on:
           Credential provider for the hermes backend: openrouter (default) or
           openai — openai-direct uses OPENAI_REVIEWER_KEY, avoids the
           OpenRouter platform fee, and takes provider-NATIVE model ids
-          (gpt-5.6-terra, no openai/ prefix).
+          (gpt-x.y, no openai/ prefix).
         type: string
         required: false
         default: openrouter

.github/workflows/advisory-review-summarize.yml

@@ -28,10 +28,10 @@ on:
         required: false
         default: ''
       backend:
-        description: Backend for the summarizer session (claude | hermes | codex).
+        description: Backend for the summarizer session (codex | claude | hermes).
         type: string
         required: false
-        default: claude
+        default: codex
       model:
         description: Model for the summarizer session (same rules as the reviewer).
         type: string

.github/workflows/maintenance-agent.yml

@@ -13,9 +13,9 @@ on:
   workflow_call:
     inputs:
       backend:
-        description: Session backend — claude, codex or hermes
+        description: Session backend — codex, claude or hermes
         type: string
-        default: claude
+        default: codex
       model:
         description: >-
           Model id for the backend (hermes+openrouter needs the provider prefix,

.github/workflows/maintenance.yml

@@ -16,9 +16,6 @@ jobs:
   digest:
     uses: outerloop-science/outerloop/.github/workflows/maintenance-agent.yml@main
     with:
-      backend: hermes
-      hermes_provider: openai
-      model: gpt-5.6-terra
       opinion_label: terra
     secrets:
       openai_reviewer_key: ${{ secrets.OPENAI_REVIEWER_KEY }}

.github/workflows/review-agent.yml

@@ -15,15 +15,15 @@ on:
       backend:
         description: Which agent backend drives the review
         type: choice
-        options: [claude, codex, hermes]
-        default: claude
+        options: [codex, claude, hermes]
+        default: codex
       model:
         description: >-
           Model for the session. The claude backend requires one: this input,
           else OUTERLOOP_CLAUDE_MODEL in the environment (no built-in default).
-          Empty -> codex configured, hermes default. For hermes pass the
+          Empty -> the kernel's pinned codex model, hermes default. For hermes pass the
           PROVIDER'S native id (openrouter: deepseek/deepseek-chat; openai:
-          gpt-5.6-terra).
+          gpt-x.y).
         type: string
         required: false
         default: ''

.github/workflows/review.yml

@@ -31,11 +31,7 @@ jobs:
     with:
       bot_login: 'outerloop-science[bot]'
       bot_aliases: 'outerloop-autoresearch[bot], agentic-learning-bot'
-      backend: hermes
-      # openai-direct: same token rate, no OpenRouter platform fee (the org's
-      # OpenAI account is tax-exempt); model id is provider-native
-      hermes_provider: openai
-      model: gpt-5.6-terra
+      # backend and model: the kernel's defaults (codex on its pinned model)
       # `general` is a real lens name meaning the full rubric (no empty-string
       # ternary — its GitHub Actions form silently falls through to the else)
       lens: ${{ matrix.lens }}
@@ -51,9 +47,6 @@ jobs:
     with:
       bot_login: 'outerloop-science[bot]'
       bot_aliases: 'outerloop-autoresearch[bot], agentic-learning-bot'
-      backend: hermes
-      hermes_provider: openai
-      model: gpt-5.6-terra
       opinion_label: terra
       # keep in lockstep with the lens matrix above
       expected_lenses: general credentials deployment lifecycle coverage prose
@@ -67,9 +60,6 @@ jobs:
     with:
       bot_login: 'outerloop-science[bot]'
       bot_aliases: 'outerloop-autoresearch[bot], agentic-learning-bot'
-      backend: hermes
-      hermes_provider: openai
-      model: gpt-5.6-terra
       opinion_label: terra
       # a maintainer's review label approves a fork PR; the fork's head is
       # never checked out, but its diff reaches the model

CHANGELOG.md

@@ -6,8 +6,13 @@ Versions follow [SemVer](https://semver.org).
 
 ## [Unreleased]
 
+### Upgrading
+
+- The reusable review, summary and maintenance workflows now default to the codex backend. A caller that relied on the old `claude` default must now set `backend: claude`; otherwise its reviews post a could-not-run stub until an `openai_reviewer_key` secret is passed.
+
 ### Changed
 
+- The reviewer, summarizer and maintenance digest default to Codex, on a model the kernel pins next to the Codex CLI version (`model` in the `[codex]` table of `harnesses.toml`). A caller that leaves `model` unset follows the pin, so moving it in a release moves every such caller. A `model` input still wins.
 - `merge: auto` merges a blessed head even while its author is still at work. Before, the merge waited until the run had no jobs in flight and no unread messages, so an author that kept revising on each panel round's advisory findings held its own cleared pull request open for days. A human's comment still holds the merge until the author has read it. Merging ends the run and cancels its jobs, as before; further work comes back as a new pull request.
 
 ## [0.6.0] - 2026-10-26

docs/reviewer.md

@@ -54,21 +54,21 @@ jobs:
       reviewer_ref: v0             # keep on the same ref as the line above
       bot_login: my-bot            # PRs by this login are never reviewed
       # personal: true             # personal mode: skip only this login's marked PRs
-      model: claude-sonnet-5-5     # required: the claude backend has no default
     secrets:
-      anthropic_reviewer_key: ${{ secrets.ANTHROPIC_REVIEWER_KEY }}
+      openai_reviewer_key: ${{ secrets.OPENAI_REVIEWER_KEY }}
 ```
 
 3. Add the key as a repository secret. Under
    `Settings → Secrets and variables → Actions`, create
-   `ANTHROPIC_REVIEWER_KEY` with your Anthropic key. A dedicated, spend-capped
+   `OPENAI_REVIEWER_KEY` with your OpenAI key. A dedicated, spend-capped
    key is strongly recommended.
 
 Open a pull request and a review appears within a minute or two. The same
 file ships as [`examples/review.yml`](https://github.com/outerloop-science/outerloop/blob/main/examples/review.yml).
-The claude backend has no built-in model, so a review workflow needs `model`
-or a repository variable `OUTERLOOP_CLAUDE_MODEL` it inherits; without one,
-every PR gets a could-not-run stub.
+The reviewer runs on Codex with the model Outerloop pins for it. A release
+that moves the pin moves every workflow on `@v0` that leaves `model` unset,
+so you get model upgrades with kernel releases. Set `model` to stay on a
+model of your own choosing.
 
 ## Re-review and opt-out
 
@@ -79,16 +79,18 @@ label opts a PR out entirely.
 
 ## Backends and second opinions
 
-`backend: claude` is the default. Two other backends work with the same
+`backend: codex` is the default. Two other backends work with the same
 workflow and their own key:
 
+- `backend: claude` uses the `ANTHROPIC_REVIEWER_KEY` secret
+  (`anthropic_reviewer_key`). It has no built-in model: set `model`, or a
+  repository variable `OUTERLOOP_CLAUDE_MODEL` the workflow inherits;
+  without one, every PR gets a could-not-run stub.
 - `backend: hermes` uses `hermes_provider: openrouter` by default with the
   `OPENROUTER_API_KEY` secret and an OpenRouter model id such as
   `openai/gpt-5.6-terra`. Add `hermes_provider: openai` to run directly
   against the OpenAI API with the `OPENAI_REVIEWER_KEY` secret and a
   provider-native id such as `gpt-5.6-terra`.
-- `backend: codex` uses the `OPENAI_REVIEWER_KEY` secret and a Codex model
-  id.
 
 Every input of the reusable workflow is documented in
 [`advisory-review-agent.yml`](https://github.com/outerloop-science/outerloop/blob/main/.github/workflows/advisory-review-agent.yml).
@@ -162,9 +164,8 @@ jobs:
     uses: outerloop-science/outerloop/.github/workflows/maintenance-agent.yml@v0
     with:
       reviewer_ref: v0
-      model: claude-sonnet-5-5   # required: the claude backend has no default
     secrets:
-      anthropic_reviewer_key: ${{ secrets.ANTHROPIC_REVIEWER_KEY }}
+      openai_reviewer_key: ${{ secrets.OPENAI_REVIEWER_KEY }}
 ```
 
 The `backend`, `model` and `hermes_provider` inputs select the model as they

examples/review.yml

@@ -9,7 +9,7 @@
 # are skipped. The model key lives only in the read-only session job.
 #
 # Setup (see docs/reviewer.md): (1) have a bot account whose PRs are skipped,
-# (2) add this file, (3) set the ANTHROPIC_REVIEWER_KEY repo secret.
+# (2) add this file, (3) set the OPENAI_REVIEWER_KEY repo secret.
 name: outerloop-review
 on:
   pull_request_target:
@@ -36,7 +36,7 @@ jobs:
       bot_login: your-bot-login
       # set it when the kernel runs with your own token (personal mode)
       # personal: true
-      backend: claude          # claude | hermes | codex
-      model: claude-sonnet-5-5   # required for claude: it has no default model
+      # backend: codex (default) | claude | hermes. Codex runs on the model
+      # this release pins unless you set `model`; claude needs `model`.
     secrets:
-      anthropic_reviewer_key: ${{ secrets.ANTHROPIC_REVIEWER_KEY }}
+      openai_reviewer_key: ${{ secrets.OPENAI_REVIEWER_KEY }}

src/outerloop/attempt.py

@@ -312,7 +312,7 @@ def author_config_error(
     if not model or model.startswith("claude"):
         return (
             "author-backend codex needs a codex/openai model in OUTERLOOP_AUTHOR_MODEL "
-            f"(e.g. gpt-5.6-terra), not a claude model or none (got {model!r})"
+            f"(e.g. gpt-x.y), not a claude model or none (got {model!r})"
         )
     return ""
 
@@ -5774,7 +5774,7 @@ def _run_id(value: str) -> str:
         help="agent backend for the author/editor role (config-driven: default "
         "from OUTERLOOP_AUTHOR_BACKEND). codex and hermes run contained (apptainer + "
         "--sandbox danger-full-access) and REQUIRES --image and a codex/openai "
-        "--model (e.g. gpt-5.6-terra).",
+        "--model (e.g. gpt-x.y).",
     )
     parser.add_argument(
         "--codex-config",

src/outerloop/harnesses.toml

@@ -9,6 +9,8 @@ linux-arm64 = "24fac77749bed3d91365d6b6915aa4b824e14318ecb6bc17adbc192f01c9173d"
 version = "0.160.0"
 sha256 = "306865417d4ee7a927785852910a527f41e1e159add390ac5ae3accb67d44a13"
 code_mode_host_sha256 = "ac6cd6288f0e39f46a33eba1cfd37711651f14f33dbc73d6a1b82eb45e038cac"
+# review sessions on codex use this model when the caller names none
+model = "gpt-5.6-terra"
 
 [hermes]
 ref = "v2026.9.24"

src/outerloop/review_agent_cli.py

@@ -15,6 +15,7 @@
 from outerloop.endpoints import resolve_endpoint
 from outerloop.github import EnvTokenProvider, GitHubClient
 from outerloop.harness import ClaudeModelUnset, Harness
+from outerloop.harness_pins import pins
 from outerloop.review_agent import (
     emit_envelope,
     run_agent_review,
@@ -73,6 +74,8 @@ def resolve_reviewer_harness(spec: RoleSpec) -> tuple[Harness | None, str, str]:
             return harness, "", backend
     except ValueError as exc:
         return None, str(exc), backend
+    if not review_model and backend == "codex":
+        review_model = pins("codex")["model"]
     hermes_provider = os.environ.get("REVIEW_HERMES_PROVIDER", "").lower() or "openrouter"
     key_var = {
         "claude": "ANTHROPIC_REVIEWER_KEY",
@@ -108,14 +111,14 @@ def resolve_reviewer_harness(spec: RoleSpec) -> tuple[Harness | None, str, str]:
             return (
                 None,
                 "hermes+openrouter requires an OpenRouter-shaped REVIEW_MODEL "
-                "(openai/gpt-5.6-terra, with the provider prefix)",
+                "(openai/gpt-x.y, with the provider prefix)",
                 backend,
             )
         if provider == "openai" and (not review_model or "/" in review_model):
             return (
                 None,
                 "hermes+openai requires a provider-native REVIEW_MODEL "
-                "(gpt-5.6-terra, no openai/ prefix)",
+                "(gpt-x.y, no openai/ prefix)",
                 backend,
             )
     try:

tests/test_review_agent_cli.py

@@ -73,6 +73,8 @@ def test_codex_backend_reads_openai_key(monkeypatch: Any) -> None:
     # the codex backend AND its openai key are forwarded to the harness builder
     assert calls["build_kwargs"]["backend"] == "codex"
     assert calls["build_key"] == "sk-openai"
+    # no REVIEW_MODEL: the kernel's pinned codex model, not codex's own default
+    assert calls["build_kwargs"]["model"] == cli.pins("codex")["model"]
 
 
 def test_hermes_backend_reads_openrouter_key(monkeypatch: Any, tmp_path: Path) -> None: