omx(team): auto-checkpoint worker-3 [4]

bellman · bellman · commit 5c40d4e778b3 · 2026-05-14T18:26:55.000+09:00
diff --git a/rust/crates/runtime/src/green_contract.rs b/rust/crates/runtime/src/green_contract.rs
@@ -30,12 +30,33 @@ impl std::fmt::Display for GreenLevel {
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 pub struct GreenContract {
     pub required_level: GreenLevel,
+    pub require_test_command_provenance: bool,
+    pub require_base_branch_freshness: bool,
+    pub require_recovery_attempt_context: bool,
+    pub block_known_flakes: bool,
 }
 
 impl GreenContract {
     #[must_use]
     pub fn new(required_level: GreenLevel) -> Self {
-        Self { required_level }
+        Self {
+            required_level,
+            require_test_command_provenance: false,
+            require_base_branch_freshness: false,
+            require_recovery_attempt_context: false,
+            block_known_flakes: false,
+        }
+    }
+
+    #[must_use]
+    pub fn merge_ready(required_level: GreenLevel) -> Self {
+        Self {
+            required_level,
+            require_test_command_provenance: true,
+            require_base_branch_freshness: true,
+            require_recovery_attempt_context: true,
+            block_known_flakes: true,
+        }
     }
 
     #[must_use]
@@ -52,12 +73,164 @@ impl GreenContract {
         }
     }
 
+    #[must_use]
+    pub fn evaluate_evidence(&self, evidence: &GreenEvidence) -> GreenEvidenceOutcome {
+        let mut missing = Vec::new();
+        let mut blocking_flakes = Vec::new();
+
+        if evidence.observed_level < self.required_level {
+            missing.push(GreenContractRequirement::RequiredLevel);
+        }
+
+        if self.require_test_command_provenance && !evidence.has_passing_test_command() {
+            missing.push(GreenContractRequirement::TestCommandProvenance);
+        }
+
+        if self.require_base_branch_freshness && !evidence.base_branch_fresh {
+            missing.push(GreenContractRequirement::BaseBranchFreshness);
+        }
+
+        if self.require_recovery_attempt_context && !evidence.recovery_attempt_context_recorded {
+            missing.push(GreenContractRequirement::RecoveryAttemptContext);
+        }
+
+        if self.block_known_flakes {
+            blocking_flakes = evidence
+                .known_flakes
+                .iter()
+                .filter(|flake| flake.blocks_green)
+                .cloned()
+                .collect();
+        }
+
+        if missing.is_empty() && blocking_flakes.is_empty() {
+            GreenEvidenceOutcome::Satisfied {
+                required_level: self.required_level,
+                observed_level: evidence.observed_level,
+            }
+        } else {
+            GreenEvidenceOutcome::Unsatisfied {
+                required_level: self.required_level,
+                observed_level: evidence.observed_level,
+                missing,
+                blocking_flakes,
+            }
+        }
+    }
+
     #[must_use]
     pub fn is_satisfied_by(self, observed_level: GreenLevel) -> bool {
         observed_level >= self.required_level
     }
 }
 
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct GreenEvidence {
+    pub observed_level: GreenLevel,
+    pub test_commands: Vec<TestCommandProvenance>,
+    pub base_branch_fresh: bool,
+    pub known_flakes: Vec<KnownFlake>,
+    pub recovery_attempt_context_recorded: bool,
+}
+
+impl GreenEvidence {
+    #[must_use]
+    pub fn new(observed_level: GreenLevel) -> Self {
+        Self {
+            observed_level,
+            test_commands: Vec::new(),
+            base_branch_fresh: false,
+            known_flakes: Vec::new(),
+            recovery_attempt_context_recorded: false,
+        }
+    }
+
+    #[must_use]
+    pub fn with_test_command(mut self, command: impl Into<String>, exit_code: i32) -> Self {
+        self.test_commands.push(TestCommandProvenance {
+            command: command.into(),
+            exit_code,
+        });
+        self
+    }
+
+    #[must_use]
+    pub fn with_base_branch_fresh(mut self, is_fresh: bool) -> Self {
+        self.base_branch_fresh = is_fresh;
+        self
+    }
+
+    #[must_use]
+    pub fn with_known_flake(mut self, test_name: impl Into<String>, blocks_green: bool) -> Self {
+        self.known_flakes.push(KnownFlake {
+            test_name: test_name.into(),
+            blocks_green,
+        });
+        self
+    }
+
+    #[must_use]
+    pub fn with_recovery_attempt_context(mut self, recorded: bool) -> Self {
+        self.recovery_attempt_context_recorded = recorded;
+        self
+    }
+
+    #[must_use]
+    pub fn has_passing_test_command(&self) -> bool {
+        self.test_commands.iter().any(TestCommandProvenance::passed)
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct TestCommandProvenance {
+    pub command: String,
+    pub exit_code: i32,
+}
+
+impl TestCommandProvenance {
+    #[must_use]
+    pub fn passed(&self) -> bool {
+        self.exit_code == 0 && !self.command.trim().is_empty()
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct KnownFlake {
+    pub test_name: String,
+    pub blocks_green: bool,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum GreenContractRequirement {
+    RequiredLevel,
+    TestCommandProvenance,
+    BaseBranchFreshness,
+    RecoveryAttemptContext,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "outcome", rename_all = "snake_case")]
+pub enum GreenEvidenceOutcome {
+    Satisfied {
+        required_level: GreenLevel,
+        observed_level: GreenLevel,
+    },
+    Unsatisfied {
+        required_level: GreenLevel,
+        observed_level: GreenLevel,
+        missing: Vec<GreenContractRequirement>,
+        blocking_flakes: Vec<KnownFlake>,
+    },
+}
+
+impl GreenEvidenceOutcome {
+    #[must_use]
+    pub fn is_satisfied(&self) -> bool {
+        matches!(self, Self::Satisfied { .. })
+    }
+}
+
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(tag = "outcome", rename_all = "snake_case")]
 pub enum GreenContractOutcome {
@@ -149,4 +322,83 @@ mod tests {
             }
         );
     }
+    #[test]
+    fn merge_ready_contract_requires_provenance_beyond_test_level() {
+        // given
+        let contract = GreenContract::merge_ready(GreenLevel::Workspace);
+        let evidence = GreenEvidence::new(GreenLevel::Workspace)
+            .with_test_command("cargo test --manifest-path rust/Cargo.toml", 0);
+
+        // when
+        let outcome = contract.evaluate_evidence(&evidence);
+
+        // then
+        assert_eq!(
+            outcome,
+            GreenEvidenceOutcome::Unsatisfied {
+                required_level: GreenLevel::Workspace,
+                observed_level: GreenLevel::Workspace,
+                missing: vec![
+                    GreenContractRequirement::BaseBranchFreshness,
+                    GreenContractRequirement::RecoveryAttemptContext,
+                ],
+                blocking_flakes: vec![],
+            }
+        );
+        assert!(!outcome.is_satisfied());
+    }
+
+    #[test]
+    fn merge_ready_contract_accepts_complete_test_provenance_context() {
+        // given
+        let contract = GreenContract::merge_ready(GreenLevel::Workspace);
+        let evidence = GreenEvidence::new(GreenLevel::MergeReady)
+            .with_test_command("cargo test --manifest-path rust/Cargo.toml", 0)
+            .with_base_branch_fresh(true)
+            .with_recovery_attempt_context(true);
+
+        // when
+        let outcome = contract.evaluate_evidence(&evidence);
+
+        // then
+        assert_eq!(
+            outcome,
+            GreenEvidenceOutcome::Satisfied {
+                required_level: GreenLevel::Workspace,
+                observed_level: GreenLevel::MergeReady,
+            }
+        );
+    }
+
+    #[test]
+    fn known_blocking_flake_prevents_green_contract_satisfaction() {
+        // given
+        let contract = GreenContract::merge_ready(GreenLevel::Workspace);
+        let evidence = GreenEvidence::new(GreenLevel::MergeReady)
+            .with_test_command("cargo test --manifest-path rust/Cargo.toml", 0)
+            .with_base_branch_fresh(true)
+            .with_recovery_attempt_context(true)
+            .with_known_flake(
+                "session_lifecycle_prefers_running_process_over_idle_shell",
+                true,
+            );
+
+        // when
+        let outcome = contract.evaluate_evidence(&evidence);
+
+        // then
+        assert_eq!(
+            outcome,
+            GreenEvidenceOutcome::Unsatisfied {
+                required_level: GreenLevel::Workspace,
+                observed_level: GreenLevel::MergeReady,
+                missing: vec![],
+                blocking_flakes: vec![KnownFlake {
+                    test_name: "session_lifecycle_prefers_running_process_over_idle_shell"
+                        .to_string(),
+                    blocks_green: true,
+                }],
+            }
+        );
+    }
 }
diff --git a/rust/crates/runtime/src/policy_engine.rs b/rust/crates/runtime/src/policy_engine.rs
@@ -58,7 +58,9 @@ impl PolicyCondition {
             Self::Or(conditions) => conditions
                 .iter()
                 .any(|condition| condition.matches(context)),
-            Self::GreenAt { level } => context.green_level >= *level,
+            Self::GreenAt { level } => {
+                context.green_contract_satisfied && context.green_level >= *level
+            }
             Self::StaleBranch => context.branch_freshness >= STALE_BRANCH_THRESHOLD,
             Self::StartupBlocked => context.blocker == LaneBlocker::Startup,
             Self::LaneCompleted => context.completed,
@@ -134,6 +136,7 @@ pub enum DiffScope {
 pub struct LaneContext {
     pub lane_id: String,
     pub green_level: GreenLevel,
+    pub green_contract_satisfied: bool,
     pub branch_freshness: Duration,
     pub blocker: LaneBlocker,
     pub review_status: ReviewStatus,
@@ -156,6 +159,7 @@ impl LaneContext {
         Self {
             lane_id: lane_id.into(),
             green_level,
+            green_contract_satisfied: false,
             branch_freshness,
             blocker,
             review_status,
@@ -171,6 +175,7 @@ impl LaneContext {
         Self {
             lane_id: lane_id.into(),
             green_level: 0,
+            green_contract_satisfied: false,
             branch_freshness: Duration::from_secs(0),
             blocker: LaneBlocker::None,
             review_status: ReviewStatus::Pending,
@@ -179,6 +184,12 @@ impl LaneContext {
             reconciled: true,
         }
     }
+
+    #[must_use]
+    pub fn with_green_contract_satisfied(mut self, satisfied: bool) -> Self {
+        self.green_contract_satisfied = satisfied;
+        self
+    }
 }
 
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -257,7 +268,8 @@ mod tests {
             ReviewStatus::Approved,
             DiffScope::Scoped,
             false,
-        );
+        )
+        .with_green_contract_satisfied(true);
 
         // when
         let actions = engine.evaluate(&context);
@@ -266,6 +278,36 @@ mod tests {
         assert_eq!(actions, vec![PolicyAction::MergeToDev]);
     }
 
+    #[test]
+    fn merge_rule_blocks_when_green_tests_lack_contract_provenance() {
+        // given
+        let engine = PolicyEngine::new(vec![PolicyRule::new(
+            "merge-to-dev",
+            PolicyCondition::And(vec![
+                PolicyCondition::GreenAt { level: 2 },
+                PolicyCondition::ScopedDiff,
+                PolicyCondition::ReviewPassed,
+            ]),
+            PolicyAction::MergeToDev,
+            20,
+        )]);
+        let context = LaneContext::new(
+            "lane-7",
+            3,
+            Duration::from_secs(5),
+            LaneBlocker::None,
+            ReviewStatus::Approved,
+            DiffScope::Scoped,
+            false,
+        );
+
+        // when
+        let actions = engine.evaluate(&context);
+
+        // then
+        assert!(actions.is_empty());
+    }
+
     #[test]
     fn stale_branch_rule_fires_at_threshold() {
         // given
@@ -468,7 +510,8 @@ mod tests {
             ReviewStatus::Pending,
             DiffScope::Full,
             false,
-        );
+        )
+        .with_green_contract_satisfied(true);
 
         // when
         let actions = engine.evaluate(&context);
diff --git a/rust/crates/runtime/tests/integration_tests.rs b/rust/crates/runtime/tests/integration_tests.rs
