feat(test-classifier): let the agent re-resolve the diff range (fix 'nothing to classify') (#75)

A user on a GitHub Enterprise host (github.cms.gov) got the PR lookup working via
AI_REVIEW_REPO, but then hit 'No change under test (git diff origin/main HEAD) —
nothing to classify'. Root cause: the dispatcher hardcodes the diff base as
origin/<base> and the prompt told the agent to 'use it verbatim'. When the PR
lives on a repo that is NOT the local origin (fork, enterprise host, or a stale/
absent origin/<base>), that range is wrong/empty — and 'verbatim' made the agent
trust the emptiness instead of recovering.

Reframe: the precomputed range is a STARTING POINT, and the agent (which has git
+ shell) re-resolves the real diff itself when it looks empty/wrong — exactly the
deterministic recovery it was already improvising in practice.

  • dispatcher exports AI_REVIEW_PR_NUMBER / AI_REVIEW_PR_BASE / AI_REVIEW_REPO_SLUG
    so the prompt can reference them; AI_REVIEW_DIFF_RANGE stays as the start.
  • prompt: drop 'use it verbatim' → 'START there, verify, do NOT treat as gospel';
    on empty/wrong range, fetch the base from the PR's OWN repo
    (gh repo view -R <slug> --json url → git fetch <url> <base>) and diff
    FETCH_HEAD HEAD. Works across remotes/hosts. Note the re-resolved base in the
    summary.
  • SKILL.md Step 2 updated to match.

Does NOT change the happy path (origin/<base> still the default when correct).
Still need a user on the fork/enterprise layout to confirm end-to-end; the change
only ADDS a recovery path, so it can't regress runs that already worked.

Author: fg-nava

testing/classifier/.skills/test-classifier/SKILL.md

@@ -196,15 +196,24 @@ marker, and stop. Passing tests are never listed.
 
 ## Step 2 — Collect the Diff
 
-The dispatcher passes the change-under-test base ref via the
-`AI_REVIEW_AGAINST` environment variable. If unset, the dispatcher will have
-refused to run.
+The dispatcher passes a **candidate** range via `AI_REVIEW_DIFF_RANGE` (and the
+base ref via `AI_REVIEW_AGAINST`). Start there:
 
 ```bash
-git diff "$AI_REVIEW_AGAINST" HEAD --unified=5      # full content of the change
-git diff "$AI_REVIEW_AGAINST" HEAD --name-only      # list of changed paths
+git diff $AI_REVIEW_DIFF_RANGE --name-only          # list of changed paths
+git diff $AI_REVIEW_DIFF_RANGE --unified=5          # full content of the change
 ```
 
+**Verify it reflects the real change under test — it is a starting point, not
+gospel.** The precomputed base (often `origin/<base>`) can be wrong or empty when
+the PR lives on a different repo than the local `origin`, on a fork or enterprise
+host, or when `origin/<base>` is stale/absent. If the name-only diff is empty or
+clearly not this PR's change, re-resolve the base yourself before concluding
+there is nothing to classify: for a PR run use `AI_REVIEW_PR_NUMBER` /
+`AI_REVIEW_REPO_SLUG` / `AI_REVIEW_PR_BASE` to fetch the base from the PR's own
+repo and diff `FETCH_HEAD HEAD`; for a local `--unpushed` run fall back to the
+merge-base with the remote default branch. Note in the summary if you re-resolved.
+
 The diff is your evidence for the central question: did this change *intend* to
 alter the behavior the failing test asserts on? An intended behavior change that
 the test wasn't updated for points toward `TEST_BUG`. A change that should NOT

testing/classifier/.skills/test-classifier/scripts/test-classifier-dispatcher.sh

@@ -401,12 +401,35 @@ in this repository at:
 
 Follow the file at the path above.
 
-Classify the failing tests for the change under test. The exact git range is in
-the AI_REVIEW_DIFF_RANGE env var (base→HEAD normally; base→index, i.e. committed
-+ staged, for a local --unpushed run). Use it verbatim:
+Classify the failing tests for the change under test. The dispatcher precomputed
+a candidate git range in the AI_REVIEW_DIFF_RANGE env var (base→HEAD normally;
+base→index for a local --unpushed run). START there:
 
-  git diff \$AI_REVIEW_DIFF_RANGE --unified=5
   git diff \$AI_REVIEW_DIFF_RANGE --name-only
+  git diff \$AI_REVIEW_DIFF_RANGE --unified=5
+
+IMPORTANT — verify the range actually reflects the change under test; do NOT
+treat AI_REVIEW_DIFF_RANGE as gospel. The precomputed base (often \`origin/<base>\`)
+can be WRONG or empty when the PR lives on a different repo than the local
+\`origin\`, on a fork, or on an enterprise host (e.g. github.cms.gov), or when
+\`origin/<base>\` is stale or absent. If \`git diff \$AI_REVIEW_DIFF_RANGE --name-only\`
+is EMPTY or obviously not this PR's change, RE-RESOLVE the base yourself before
+concluding there is nothing to classify:
+
+  • This is a PR run when AI_REVIEW_PR_NUMBER is set; its repo is in
+    AI_REVIEW_REPO_SLUG and its base branch in AI_REVIEW_PR_BASE.
+  • Fetch the PR's base from its own repo and diff HEAD against what you fetched:
+      gh pr view "\$AI_REVIEW_PR_NUMBER" -R "\$AI_REVIEW_REPO_SLUG" --json baseRefName,headRefOid
+      # fetch the base ref from the PR's repo (works across remotes/hosts):
+      url="\$(gh repo view -R "\$AI_REVIEW_REPO_SLUG" --json url --jq .url)"
+      git fetch "\$url" "\$AI_REVIEW_PR_BASE"
+      git diff FETCH_HEAD HEAD --name-only
+      git diff FETCH_HEAD HEAD --unified=5
+  • For a local --unpushed run (no PR number), fall back to the merge-base with
+    the remote default branch, or the last pushed commit, as the base.
+
+Use whichever range genuinely captures the change under test. State in the JSON
+"summary" which base you used if you had to re-resolve it.
 
 Follow the skill instructions in that SKILL.md file (path above) exactly:
 
@@ -947,10 +970,18 @@ test_classifier::run() {
   ai_review::log  "────────────────────────────────────────────────────────────"
 
   export AI_REVIEW_AGAINST
+  # Export the PR context so the agent can re-resolve the diff itself when the
+  # precomputed range below is empty/wrong (forks, enterprise hosts, stale
+  # origin/<base>). These may be empty for a local --unpushed run.
+  export AI_REVIEW_PR_NUMBER="${AI_REVIEW_PR_NUMBER:-}"
+  export AI_REVIEW_PR_BASE="${AI_REVIEW_PR_BASE:-}"
+  export AI_REVIEW_REPO_SLUG="${AI_REVIEW_REPO_SLUG:-}"
 
   # The git range the AI should diff, matching the dispatcher's own accounting:
   #   --unpushed (INCLUDE_STAGED) → `--cached <base>`  (committed + staged)
   #   otherwise                   → `<base> HEAD`      (committed range)
+  # This is a STARTING POINT, not gospel — the prompt authorizes the agent to
+  # re-resolve locally if it doesn't reflect the real change under test.
   if [[ "${AI_REVIEW_INCLUDE_STAGED:-0}" == "1" ]]; then
     AI_REVIEW_DIFF_RANGE="--cached ${AI_REVIEW_AGAINST}"
   else