Require proof before superseded PR closeout

.github/workflows/sweep.yml

@@ -1815,6 +1815,13 @@ jobs:
         with:
           build-script: build:all
 
+      - uses: ./.github/actions/setup-codex
+        if: ${{ !((github.event_name == 'repository_dispatch' && github.event.action == 'clawsweeper_hatch') || (github.event_name == 'workflow_dispatch' && github.event.inputs.apply_sync_comments_only == 'true') || (github.event_name == 'schedule' && github.event.schedule == '6,21,36,51 * * * *')) }}
+        env:
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+        with:
+          login-status: "true"
+
       - name: Sync before applying decisions
         run: git pull --rebase
 

prompts/repair/replacement-closeout-proof.md

@@ -0,0 +1,17 @@
+You are ClawSweeper's read-only replacement closeout proof checker.
+
+Decide whether PR B can safely close PR A as superseded.
+
+Hard rules:
+- You only have two decisions: `superseded` or `keep_open`.
+- PR B may be user-authored and may have a different author from PR A.
+- A source list or `supersedes #A` text is only a candidate signal.
+- Compare the useful work generally from the compact context: title, first body excerpt, labels, file paths, file counts, timestamps, and repair provenance.
+- Do not require exact patch-line equality. A replacement can cover the same behavior with different code shape.
+- Return `superseded` only when PR B clearly covers PR A's useful work and PR A has no unique behavior, file concern, proof, discussion, or review point needing separate maintainer review.
+- Return `keep_open` for anything else, including related PRs, incomplete proof, thin context, or uncertainty.
+- Security-sensitive work is not a separate close blocker. Treat security labels, CVE/GHSA text, and ClawSweeper security markers as PR A content to compare. If PR B proves it covers that content, PR A can be `superseded`.
+- Use `securityBlocked: true` only when PR A has security-sensitive content that PR B does not prove it covers. Also list that uncovered content in `uniqueSourceWork`.
+- Do not ask for more context.
+
+Return only JSON matching the supplied schema.

prompts/review-item.md

@@ -169,6 +169,14 @@ discussion:
 `merge-risk: 🚨 security-boundary`: 🚨 Merging this PR could weaken sandboxing, authorization, credentials, or sensitive data.
 `merge-risk: 🚨 availability`: 🚨 Merging this PR could cause crashes, hangs, restart loops, stalls, or process outages.
 `merge-risk: 🚨 automation`: 🚨 Merging this PR could break CI, automerge, proof capture, label sync, or automation.
+Do not treat a branch being behind the current base as proof that merging the
+PR will delete current-base-only files or commits. When GitHub reports the PR as
+mergeable or clean and the only concern is stale base drift, describe it as
+needing rebase or review refresh in `risks`, `workReason`, or `bestSolution`,
+but leave `reviewFindings` and `mergeRiskLabels` focused on defects or risks
+that survive the actual three-way merge result. Use deletion/drop wording for
+current-base behavior only when a merge result, merge ref, conflict, or concrete
+patch evidence shows that the merged PR would remove or regress it.
 When merge risk is present, explain it in `risks` in maintainer-facing language
 and make `bestSolution` the best end state. Fill `mergeRiskOptions` with 1-3
 risk-specific maintainer options. Do not use a fixed menu. Each option needs a

prompts/supersession-proof.md

@@ -0,0 +1,19 @@
+You are ClawSweeper's read-only supersession proof checker.
+
+Decide whether PR B can safely supersede PR A.
+
+Hard rules:
+- You only have two decisions: `superseded` or `keep_open`.
+- PR B may be user-authored and may have a different author from PR A.
+- Text such as `supersedes #A` is only a candidate signal.
+- Compare the useful work generally from the compact context: title, first body excerpt, labels, file paths, file counts, and timestamps.
+- Do not require exact patch-line equality. A replacement can cover the same behavior with different code shape.
+- Return `superseded` only when PR B clearly covers PR A's useful work and PR A has no unique behavior, file concern, proof, discussion, or review point needing separate maintainer review.
+- Return `keep_open` for anything else, including related PRs, incomplete proof, thin context, or uncertainty.
+- Security-sensitive work is not a separate close blocker. Treat security labels, CVE/GHSA text, and ClawSweeper security markers as PR A content to compare. If PR B proves it covers that content, PR A can be `superseded`.
+- Use `securityBlocked: true` only when PR A has security-sensitive content that PR B does not prove it covers. Also list that uncovered content in `uniqueSourceWork`.
+- `coveredWork` must describe concrete PR A work that PR B covers.
+- `uniqueSourceWork` must list any PR A behavior, file concern, proof, discussion, or review point that remains unique. Use an empty array only when none remains.
+- Do not ask for more context.
+
+Return only JSON matching the supplied schema.

schema/clawsweeper-supersession-proof.schema.json

@@ -0,0 +1,46 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "sourceSummary",
+    "replacementSummary",
+    "coveredWork",
+    "uniqueSourceWork",
+    "securityBlocked",
+    "decision",
+    "reason"
+  ],
+  "properties": {
+    "sourceSummary": {
+      "type": "string",
+      "description": "Concrete summary of the useful work in PR A."
+    },
+    "replacementSummary": {
+      "type": "string",
+      "description": "Concrete summary of the useful work in PR B."
+    },
+    "coveredWork": {
+      "type": "array",
+      "items": { "type": "string" },
+      "description": "Concrete PR A behavior, file concerns, proof, or review points that PR B covers."
+    },
+    "uniqueSourceWork": {
+      "type": "array",
+      "items": { "type": "string" },
+      "description": "Any PR A work that PR B does not cover and still needs separate review."
+    },
+    "securityBlocked": {
+      "type": "boolean",
+      "description": "True only when PR A has security-sensitive content that PR B does not prove it covers. Security-sensitive content is not a standalone close blocker when coverage is proven."
+    },
+    "decision": {
+      "type": "string",
+      "enum": ["superseded", "keep_open"]
+    },
+    "reason": {
+      "type": "string",
+      "description": "Short explanation for the decision."
+    }
+  }
+}