Keep hosted scorer routing available

.github/workflows/release.yml

@@ -113,6 +113,25 @@ jobs:
         working-directory: bench
         run: bun install --frozen-lockfile
 
+      - name: Run the embedded hosted scorer gate
+        working-directory: bench
+        env:
+          OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
+          POSTIL_SCORER_EVAL_REPEATS: "3"
+          POSTIL_SCORER_EVAL_UPSTREAM_PROVIDER: Azure
+        run: >-
+          bun run scorer-eval --json-out
+          "${{ runner.temp }}/scorer-eval-report.json"
+
+      - name: Upload the scorer gate report
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: scorer-eval-report
+          path: ${{ runner.temp }}/scorer-eval-report.json
+          if-no-files-found: warn
+          retention-days: 30
+
       - name: Run diff-file live benchmark
         working-directory: bench
         env:

README.md

@@ -35,7 +35,7 @@ postil hook install                      # add a pre-push review
 
 Bare local review selects staged changes first, then committed changes since the current branch's locally known default branch, then tracked working-tree changes, and finally an empty clean diff. Symbolic remote HEAD plus conventional `main`, `master`, and `trunk` refs are recognized without fetching. If no default branch can be resolved for a repository with commits, Postil fails closed and asks for `--base`, `--staged`, or `--diff-file` instead of reporting a partial review as clean. It sends the selected diff to the configured inference endpoint, but does not modify the working tree, index, or refs and does not write comments or checks to a forge unless `--publish` is supplied.
 
-`postil models` answers model-selection questions offline without reading credentials or contacting a provider. No model setting is required: it identifies the embedded default, explains the model-ID contracts for OpenAI-compatible and native Anthropic endpoints, separates local protocol compatibility from hosted qualification, and gives exact `postil doctor` and override commands. The Luna reviewer uses `low` reasoning effort. Local scoring stays disabled until `REVIEW_SCORER_MODEL` names a scorer; the hosted scorer candidate uses `none`. `postil review` exits `0` when the gate passes, `1` when it fails, and `2` when it cannot produce a review envelope. Interactive terminals show concise progress; `--no-progress` or `POSTIL_NO_PROGRESS=1` replaces animation with concise static milestones, while `--verbose` enables complete provider, retry, and batch telemetry.
+`postil models` answers model-selection questions offline without reading credentials or contacting a provider. No model setting is required: it identifies the embedded default, explains the model-ID contracts for OpenAI-compatible and native Anthropic endpoints, separates local protocol compatibility from hosted qualification, and gives exact `postil doctor` and override commands. The Luna reviewer and hosted scorer candidate use `low` reasoning effort. Local scoring stays disabled until `REVIEW_SCORER_MODEL` names a scorer. `postil review` exits `0` when the gate passes, `1` when it fails, and `2` when it cannot produce a review envelope. Interactive terminals show concise progress; `--no-progress` or `POSTIL_NO_PROGRESS=1` replaces animation with concise static milestones, while `--verbose` enables complete provider, retry, and batch telemetry.
 
 ## Review a pull request
 
@@ -66,7 +66,7 @@ gate:
   onError: block
 model:
   reasoningEffort: low
-  scorerReasoningEffort: none
+  scorerReasoningEffort: low
 ```
 
 Unknown keys are rejected. Repository configuration cannot redirect a deployment credential to another API host unless the operator explicitly permits that behavior. The embedded local preset uses `openai/gpt-5.6-luna`; local scoring is disabled until explicitly configured. Hosted profile selection remains service-controlled.

bench/src/livemodels.test.ts

@@ -1339,6 +1339,16 @@ describe("managed admission workflow", () => {
     // still benchmarked the previous one and passed against its old baseline.
     expect(release).not.toContain("REVIEW_MODEL:");
     expect(release).toContain("OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}");
+    expect(release).not.toContain("POSTIL_SCORER_EVAL_MODELS:");
+    expect(release).toContain('POSTIL_SCORER_EVAL_REPEATS: "3"');
+    expect(release).toContain("POSTIL_SCORER_EVAL_UPSTREAM_PROVIDER: Azure");
+    expect(release).toContain("bun run scorer-eval --json-out");
+    expect(release).toMatch(
+      /name: Upload the scorer gate report\n\s+if: always\(\)[\s\S]*if-no-files-found: warn[\s\S]*retention-days: 30/u,
+    );
+    expect(release.indexOf("bun run scorer-eval --json-out")).toBeLessThan(
+      release.indexOf("bun run bench:live --"),
+    );
     expect(release).toContain("bun run bench:live --");
     expect(release).toContain("bun run bench:compare --");
     expect(release).toMatch(/build:\n\s+needs: \[validate-tag, bench-live\]/u);

bench/src/verify-admission.test.ts

@@ -64,9 +64,14 @@ describe("admission attestation verification", () => {
   test("binds reviewer and scorer reasoning efforts to the model-defaults digest", async () => {
     const repositoryRoot = join(import.meta.dir, "..", "..");
     const original = await readFile(join(repositoryRoot, "config.toml"), "utf8");
+    const alteredScorer = original.replace(
+      /(\[scorer\][\s\S]*?reasoning_effort = )"low"/u,
+      '$1"none"',
+    );
+    expect(alteredScorer).not.toBe(original);
     for (const altered of [
       original.replace('reasoning_effort = "low"', 'reasoning_effort = "high"'),
-      original.replace('reasoning_effort = "none"', 'reasoning_effort = "low"'),
+      alteredScorer,
     ]) {
       const directory = await temporaryDirectory();
       const manifest = join(directory, "qualified-models.json");

config.toml

@@ -9,6 +9,6 @@ api_format = "openai-compatible"
 [scorer]
 enabled = true
 default_model = "openai/gpt-5.6-luna"
-reasoning_effort = "none"
+reasoning_effort = "low"
 fallback = ""
 qualification_candidates = ["openai/gpt-5.6-luna"]

docs/configuration.md

@@ -38,7 +38,7 @@ model:
   reasoningEffort: low           # max, xhigh, high, medium, low, minimal, or none
   cascade: []                    # qualified fallbacks only
   # scorer: provider/qualified-scorer
-  scorerReasoningEffort: none
+  scorerReasoningEffort: low
   consensus: 1
 ```
 
@@ -97,4 +97,4 @@ postil doctor
 
 `postil config` prints the resolved non-secret configuration and separate provenance for the model, reviewer reasoning effort, and scorer reasoning effort. `postil doctor` validates endpoint reachability, credential acceptance, and repository setup without printing credential values. Both commands identify renewable logins, access expiry, refresh inactivity expiry, and legacy access-only logins.
 
-Use `--reasoning-effort` and `--scorer-reasoning-effort` for one review. These flags override the matching environment variables, which override `model.reasoningEffort` and `model.scorerReasoningEffort`. The built-in reviewer and scorer defaults are `low` and `none`, respectively. Every request carries the resolved value, including retries and repair calls.
+Use `--reasoning-effort` and `--scorer-reasoning-effort` for one review. These flags override the matching environment variables, which override `model.reasoningEffort` and `model.scorerReasoningEffort`. The built-in reviewer and scorer defaults are both `low`. Every request carries the resolved value, including retries and repair calls.

docs/model-providers.md

@@ -4,7 +4,7 @@ Postil speaks either an OpenAI-compatible chat-completions interface or the nati
 
 ## Model admission
 
-Local review defaults to the embedded `openai/gpt-5.6-luna` reviewer with no fallback chain or scorer, so no model setting is required. Reviewer requests use `low` reasoning effort. Local scoring is disabled unless `REVIEW_SCORER_MODEL` names a scorer. The embedded hosted scorer candidate is `openai/gpt-5.6-luna` at `none` effort. `postil models` reports these roles, the accepted endpoint model-ID contracts, embedded qualification status, accepted effort values, and exact `postil doctor` and override commands without reading credentials or contacting a provider. Local use has no fixed model-ID allowlist: Postil passes a non-empty model ID unchanged to the configured OpenAI-compatible or native Anthropic endpoint. Endpoint protocol compatibility does not imply hosted qualification. Hosted deployments admit only complete qualified profiles. `REVIEW_SCORER_MODEL_CASCADE` accepts one qualified scorer fallback.
+Local review defaults to the embedded `openai/gpt-5.6-luna` reviewer with no fallback chain or scorer, so no model setting is required. Reviewer requests use `low` reasoning effort. Local scoring is disabled unless `REVIEW_SCORER_MODEL` names a scorer. The embedded hosted scorer candidate is `openai/gpt-5.6-luna` at `low` effort. `postil models` reports these roles, the accepted endpoint model-ID contracts, embedded qualification status, accepted effort values, and exact `postil doctor` and override commands without reading credentials or contacting a provider. Local use has no fixed model-ID allowlist: Postil passes a non-empty model ID unchanged to the configured OpenAI-compatible or native Anthropic endpoint. Endpoint protocol compatibility does not imply hosted qualification. Hosted deployments admit only complete qualified profiles. `REVIEW_SCORER_MODEL_CASCADE` accepts one qualified scorer fallback.
 
 Hosted deployments admit only complete profiles listed in the embedded `qualified-models.json` qualification manifest. A profile binds the canonical API base, provider API format, ordered generator chain, consensus width, ordered scorer chain, exact Bun evaluator runtime, qualification source commit, a 30-day authority window, at least three complete runs, and SHA-256 digests for the review contract, evaluator contract, fixture set, and benchmark report. The manifest also binds the exact `config.toml` digest, including the reviewer and scorer reasoning defaults. Hosted configuration must match one unexpired profile exactly, so an effort override that differs from those qualified defaults is rejected. The expiry second is excluded. An empty, degraded, stale, or mismatched profile rejects inference.
 
@@ -33,7 +33,7 @@ postil doctor
 
 These requests authenticate with `Authorization: Bearer`.
 
-OpenAI-compatible requests carry the [OpenRouter reasoning configuration](https://openrouter.ai/docs/guides/best-practices/reasoning-tokens) as `reasoning.effort`. The accepted values are `max`, `xhigh`, `high`, `medium`, `low`, `minimal`, and `none`. This field is sent to OpenRouter and custom OpenAI-compatible endpoints on every request. Postil does not drop or downgrade it, so custom endpoints must implement the extension.
+OpenAI-compatible requests carry the [OpenRouter reasoning configuration](https://openrouter.ai/docs/guides/best-practices/reasoning-tokens) as `reasoning.effort`. The accepted values are `max`, `xhigh`, `high`, `medium`, `low`, `minimal`, and `none`. Scorer requests exclude returned reasoning unless reasoning is disabled. Postil does not drop or downgrade the selected effort, so custom endpoints must implement the extension.
 
 ## Anthropic Messages API
 

qualified-models.json

@@ -4,6 +4,6 @@
   "qualificationIssuedAtUnixSeconds": null,
   "qualificationExpiresAtUnixSeconds": null,
   "qualificationMaxAgeDays": null,
-  "modelDefaultsSha256": "b581d37af77c48552412c383cebcab55129c1d8c73d3f6dfa00bf3c7ff04ab3f",
+  "modelDefaultsSha256": "535d68cf62ce90c0fa11451a5411f7f358ba73b989c1559a1374357c1384697f",
   "profiles": []
 }

src/cli.rs

@@ -141,7 +141,7 @@ pub enum Command {
         /// Reviewer reasoning effort: max|xhigh|high|medium|low|minimal|none (else REVIEW_REASONING_EFFORT, else config, else low).
         #[arg(long, value_name = "EFFORT")]
         reasoning_effort: Option<String>,
-        /// Scorer reasoning effort: max|xhigh|high|medium|low|minimal|none (else REVIEW_SCORER_REASONING_EFFORT, else config, else none).
+        /// Scorer reasoning effort: max|xhigh|high|medium|low|minimal|none (else REVIEW_SCORER_REASONING_EFFORT, else config, else low).
         #[arg(long, value_name = "EFFORT")]
         scorer_reasoning_effort: Option<String>,
         /// Keep detailed provider, retry, and batch telemetry in interactive terminals.

src/config.rs

@@ -2766,7 +2766,7 @@ scorer = { enabled = true, default_model = "provider/scorer", reasoning_effort =
         assert_eq!(c.model_source, "embedded default");
         assert_eq!(c.reasoning_effort, ReasoningEffort::Low);
         assert_eq!(c.reasoning_effort_source, "embedded default");
-        assert_eq!(c.scorer_reasoning_effort, ReasoningEffort::None);
+        assert_eq!(c.scorer_reasoning_effort, ReasoningEffort::Low);
         assert_eq!(c.scorer_reasoning_effort_source, "embedded default");
         assert!(c.concise_findings);
         assert!(!c.scorer_enabled);
@@ -3058,7 +3058,7 @@ scorer = { enabled = true, default_model = "provider/scorer", reasoning_effort =
         assert_eq!(c.model, model_defaults().default_model);
         assert_eq!(c.cascade, model_defaults().cascade);
         assert_eq!(c.reasoning_effort, ReasoningEffort::Low);
-        assert_eq!(c.scorer_reasoning_effort, ReasoningEffort::None);
+        assert_eq!(c.scorer_reasoning_effort, ReasoningEffort::Low);
         assert!(c.scorer.is_empty());
     }
 
@@ -3694,6 +3694,7 @@ scorer = { enabled = true, default_model = "provider/scorer", reasoning_effort =
         assert!(defaults.cascade.is_empty());
         assert!(defaults.scorer_enabled);
         assert_eq!(defaults.scorer_model, "openai/gpt-5.6-luna");
+        assert_eq!(defaults.scorer_reasoning_effort, ReasoningEffort::Low);
         assert!(defaults.scorer_fallback.is_empty());
         assert_eq!(
             defaults.scorer_qualification_candidates,

src/llm.rs

@@ -5277,7 +5277,11 @@ fn apply_openrouter_scorer_contract(body: &mut serde_json::Value, expected_len:
         .as_object_mut()
         .expect("provider routing configuration is an object")
         .insert("require_parameters".to_string(), json!(true));
-    body["reasoning"]["exclude"] = json!(true);
+    // A disabled reasoning request has no reasoning payload to exclude. Keep
+    // that provider contract minimal instead of sending redundant controls.
+    if body["reasoning"]["effort"] != "none" {
+        body["reasoning"]["exclude"] = json!(true);
+    }
     body["response_format"] = json!({
         "type": "json_schema",
         "json_schema": {
@@ -9555,7 +9559,7 @@ mod tests {
         );
         assert_eq!(
             scorer["reasoning"],
-            json!({"effort": "none", "exclude": true})
+            json!({"effort": "low", "exclude": true})
         );
         assert!(scorer["reasoning"].get("enabled").is_none());
         assert_eq!(scorer["provider"]["require_parameters"], true);
@@ -9611,6 +9615,27 @@ mod tests {
         assert_eq!(generator["reasoning"], json!({"effort": "low"}));
         assert!(generator.get("response_format").is_none());
         assert!(generator["provider"].get("require_parameters").is_none());
+
+        let no_reasoning_client = LlmClient::build(
+            &Config {
+                scorer_reasoning_effort: ReasoningEffort::None,
+                ..Config::default()
+            },
+            "test-key".into(),
+            Duration::from_secs(1),
+            None,
+            None,
+        )
+        .unwrap();
+        let no_reasoning_scorer = no_reasoning_client.request_body(
+            "provider/scorer",
+            "system",
+            "user",
+            400,
+            0.0,
+            LlmPhase::Scorer { expected_len: 1 },
+        );
+        assert_eq!(no_reasoning_scorer["reasoning"], json!({"effort": "none"}));
     }
 
     #[test]
@@ -10166,7 +10191,7 @@ mod tests {
             LlmPhase::Scorer { expected_len: 1 },
         );
         assert!(openai_body.get("provider").is_none());
-        assert_eq!(openai_body["reasoning"], json!({"effort": "none"}));
+        assert_eq!(openai_body["reasoning"], json!({"effort": "low"}));
         assert!(openai_body.get("response_format").is_none());
 
         let anthropic = LlmClient::build(
@@ -10191,9 +10216,9 @@ mod tests {
             LlmPhase::Scorer { expected_len: 1 },
         );
         assert!(anthropic_body.get("provider").is_none());
-        assert_eq!(anthropic_body["thinking"], json!({"type": "disabled"}));
-        assert!(anthropic_body.get("output_config").is_none());
-        assert_eq!(anthropic_body["temperature"], 0.0);
+        assert!(anthropic_body.get("thinking").is_none());
+        assert_eq!(anthropic_body["output_config"], json!({"effort": "low"}));
+        assert!(anthropic_body.get("temperature").is_none());
         assert!(anthropic_body.get("response_format").is_none());
         assert_eq!(anthropic_body["system"], "system");
     }

tests/e2e.rs

@@ -2150,7 +2150,7 @@ async fn native_anthropic_findings_use_explicit_native_scorer() {
         .body_json()
         .unwrap();
     assert_eq!(generator["output_config"], json!({"effort": "low"}));
-    assert_eq!(scorer["thinking"], json!({"type": "disabled"}));
+    assert_eq!(scorer["output_config"], json!({"effort": "low"}));
 }
 
 #[tokio::test]
@@ -6037,7 +6037,7 @@ fn hosted_config_ignores_repository_model_provider_and_scorer() {
     assert!(stdout.contains("model.scorer: "));
     assert!(!stdout.contains("model.scorer: openai/gpt-5.6-luna"));
     assert!(stdout.contains("model.reasoningEffort: low"));
-    assert!(stdout.contains("model.scorerReasoningEffort: none"));
+    assert!(stdout.contains("model.scorerReasoningEffort: low"));
     assert!(stdout.contains("model.apiBase: https://openrouter.ai/api/v1"));
     assert!(stdout.contains("model.apiFormat: openai-compatible"));
     assert!(stdout.contains("model.consensus: 1"));
@@ -6071,7 +6071,7 @@ fn provisional_hosted_config_uses_only_the_baked_roster() {
     assert!(stdout.contains("model.cascade: []"));
     assert!(stdout.contains("model.scorer: openai/gpt-5.6-luna"));
     assert!(stdout.contains("model.reasoningEffort: low"));
-    assert!(stdout.contains("model.scorerReasoningEffort: none"));
+    assert!(stdout.contains("model.scorerReasoningEffort: low"));
     assert!(stdout.contains("model.apiBase: https://openrouter.ai:443/api/v1"));
     assert!(stdout.contains("model.apiFormat: openai-compatible"));
     assert!(stdout.contains("model.consensus: 1"));
@@ -6129,7 +6129,7 @@ async fn local_review_reports_grounded_finding_and_gates() {
     let adjudication: Value = requests[1].body_json().unwrap();
     assert_eq!(request["max_tokens"], 8_000);
     assert_eq!(request["reasoning"], json!({"effort": "low"}));
-    assert_eq!(adjudication["reasoning"], json!({"effort": "none"}));
+    assert_eq!(adjudication["reasoning"], json!({"effort": "low"}));
     assert_eq!(request["messages"].as_array().unwrap().len(), 2);
     let prompt_bytes = request["messages"]
         .as_array()
@@ -6634,7 +6634,7 @@ fn models_and_config_explain_the_embedded_default() {
     assert!(models.contains("Reviewer reasoning effort: low"));
     assert!(models.contains("Local scorer: disabled"));
     assert!(models.contains("Hosted scorer candidate: openai/gpt-5.6-luna"));
-    assert!(models.contains("Hosted scorer reasoning effort: none"));
+    assert!(models.contains("Hosted scorer reasoning effort: low"));
     assert!(models.contains("does not maintain a fixed local model-ID allowlist"));
     assert!(models.contains("OpenAI-compatible endpoints accept any non-empty endpoint model ID"));
     assert!(models.contains("OpenRouter commonly uses provider/model"));
@@ -6665,7 +6665,7 @@ fn models_and_config_explain_the_embedded_default() {
     assert!(config.contains("model.reasoningEffort: low"));
     assert!(config.contains("model.reasoningEffort.source: embedded default"));
     assert!(config.contains("model.scorer.enabled: false"));
-    assert!(config.contains("model.scorerReasoningEffort: none"));
+    assert!(config.contains("model.scorerReasoningEffort: low"));
     assert!(config.contains("model.scorerReasoningEffort.source: embedded default"));
 }