fix(review): keep reviews 14 days and base graphs 30

README.md

@@ -88,7 +88,7 @@ Every bundle the action publishes carries a `metadata.json` with a `kind` of
 `review`, `base` or `warmstart`. Check it rather than inferring from the
 payload: a base bundle is otherwise indistinguishable from a head one.
 
-The action requests 30-day retention; a repository or organisation policy can shorten it, so treat an artifact's own `expired` flag as the truth rather than any fixed window.
+The action requests 14-day retention for it; a repository or organisation policy can shorten that, so treat an artifact's own `expired` flag as the truth rather than any fixed window. When a pull request outlives its review artifact, `/codeboarding` regenerates it at the cost of one incremental over the pull request — the base graph is kept longer and does not need re-analyzing.
 
 ### Reused analysis
 

action.yml

@@ -222,7 +222,9 @@ runs:
         REPOSITORY: ${{ github.repository }}
         ARTIFACT_NAME: ${{ steps.state.outputs.base_name }}
         DEST: ${{ runner.temp }}/cb-state/${{ github.action }}/base
-        RENEW_WITHIN_DAYS: '30'
+        # Must match the review retention below: a review references a base by
+        # id for its whole life, so a base with less than that left is renewed.
+        RENEW_WITHIN_DAYS: '14'
       run: "$GITHUB_ACTION_PATH/scripts/action/fetch-state.sh"
 
     - name: Fetch this pull request's last analysis
@@ -283,14 +285,13 @@ runs:
       with:
         name: codeboarding-base-${{ steps.state.outputs.cfg_hash }}-${{ steps.sync_commit.outputs.baseline_sha }}
         path: ${{ runner.temp }}/cb-state/${{ github.action }}/out/base
-        # Longer than a review's 30 days on purpose, and the margin is the whole
+        # Longer than a review's 14 days on purpose, and the margin is the whole
         # point: correctness comes from the renewal check, which republishes any
         # base with less than a review's lifetime left, while this surplus
         # decides how often that fires. Equal lifetimes renew on every run and
         # publish a base per run rather than per merge base — more storage, not
-        # less. 60 leaves a 30-day reuse window, so a pull request open a month
-        # never triggers one.
-        retention-days: 60
+        # less. 30 leaves a 16-day reuse window.
+        retention-days: 30
         if-no-files-found: ignore
 
     - name: Publish baseline analysis for the analyzed commit
@@ -300,14 +301,13 @@ runs:
       with:
         name: codeboarding-base-${{ steps.state.outputs.cfg_hash }}-${{ steps.sync_commit.outputs.analyzed_sha }}
         path: ${{ runner.temp }}/cb-state/${{ github.action }}/out/base
-        # Longer than a review's 30 days on purpose, and the margin is the whole
+        # Longer than a review's 14 days on purpose, and the margin is the whole
         # point: correctness comes from the renewal check, which republishes any
         # base with less than a review's lifetime left, while this surplus
         # decides how often that fires. Equal lifetimes renew on every run and
         # publish a base per run rather than per merge base — more storage, not
-        # less. 60 leaves a 30-day reuse window, so a pull request open a month
-        # never triggers one.
-        retention-days: 60
+        # less. 30 leaves a 16-day reuse window.
+        retention-days: 30
         if-no-files-found: ignore
 
     - name: Write sync summary
@@ -371,14 +371,13 @@ runs:
       with:
         name: ${{ steps.state.outputs.base_name }}
         path: ${{ runner.temp }}/cb-state/${{ github.action }}/out/base
-        # Longer than a review's 30 days on purpose, and the margin is the whole
+        # Longer than a review's 14 days on purpose, and the margin is the whole
         # point: correctness comes from the renewal check, which republishes any
         # base with less than a review's lifetime left, while this surplus
         # decides how often that fires. Equal lifetimes renew on every run and
         # publish a base per run rather than per merge base — more storage, not
-        # less. 60 leaves a 30-day reuse window, so a pull request open a month
-        # never triggers one.
-        retention-days: 60
+        # less. 30 leaves a 16-day reuse window.
+        retention-days: 30
         if-no-files-found: ignore
 
     - name: Render review diagram
@@ -419,9 +418,11 @@ runs:
         name: codeboarding-review-${{ github.run_id }}-${{ github.run_attempt }}
         path: ${{ steps.review_artifact.outputs.artifact_dir }}
         if-no-files-found: error
-        # Each run carries its own copy of the base graph, and the webview only
-        # ever reads a pull request's latest artifact, so value drops off fast.
-        retention-days: 30
+        # The dominant storage cost, since it is the one kept for weeks. A pull
+        # request open longer than this loses its rendered analysis until someone
+        # asks for it again, which costs one incremental over the pull request:
+        # the base graph is still published and needs no re-analysis.
+        retention-days: 14
 
     - name: Build review comment
       id: review_body

docs/COMMIT_STRATEGY.md

@@ -27,7 +27,7 @@ cache was free. Retention is the lever — see below.
 
 | Artifact | Contents | Retention | Read by |
 |---|---|---|---|
-| `codeboarding-review-<run>-<attempt>` | `analysis.json`, `health_report.json`, `metadata.json` | 30 days | the webview, humans |
+| `codeboarding-review-<run>-<attempt>` | `analysis.json`, `health_report.json`, `metadata.json` | 14 days | the webview, humans |
 | `codeboarding-base-<cfg>-<merge_base>` | the merge base's own analysis | 30 days, renewed while still referenced | any later review forking from that commit |
 | `codeboarding-warmstart-<cfg>-pr<N>` | the working directory: graph, pickle, fingerprint, gate | **1 day**, configurable | only the next run of that pull request |
 
@@ -71,9 +71,13 @@ that commit has a different merge base.
 Artifacts are charged by size × time, so the three windows are set by what reads
 them:
 
-- **30 days** for the review artifact — a reader may come back to a pull request.
-- **Repository default** for a base graph — it must outlive every review artifact
-  that names it.
+- **14 days** for the review artifact — the dominant cost, since it is the one
+  kept for weeks. A pull request open longer loses its rendered analysis until
+  someone asks for it again, which costs one incremental over the pull request:
+  the base graph is still published, so nothing re-analyzes the base.
+- **30 days** for a base graph, which must outlive every review that names it.
+  The renewal threshold is the review's own retention, so the surplus — 16 days
+  here — is how long a base is reused before being republished.
 - **1 day** for the warm-start bundle, since only the next run reads it. This is
   `warmstart_retention_days` if a repository wants longer. It behaves like the
   old cache eviction: a pull request left alone longer than the window

tests/test_action_state.py

@@ -497,6 +497,15 @@ def test_a_base_outlives_the_reviews_that_reference_it(self) -> None:
                 f"{step['name']} does not outlive the reviews that reference it",
             )
 
+    def test_the_renewal_threshold_matches_the_review_retention(self) -> None:
+        # A review references a base by id for its whole life, so the threshold
+        # that triggers renewal has to be that same life. If the two drift apart,
+        # a review can outlive the base it names and nothing catches it.
+        text = (ROOT / "action.yml").read_text(encoding="utf-8")
+        renew = next(l for l in text.splitlines() if "RENEW_WITHIN_DAYS:" in l)
+        review = next(s for s in self._uploads() if "review_artifact.outputs.artifact_dir" in s.get("path", ""))
+        self.assertIn(f"'{review['retention-days']}'", renew)
+
     def test_the_reusable_analysis_honours_the_configured_retention(self) -> None:
         warmstart = [s for s in self._uploads() if "warmstart" in s.get("path", "")]
         self.assertTrue(warmstart, "no warm-start publication step found")