fix(code-review): base local diff on the branch fork point, not a stale ref

CHANGELOG.md

@@ -4,6 +4,13 @@ All notable changes to the claude-plugins project will be documented in this fil
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Entries are listed newest-first; each plugin section is treated as released when merged to `main`.
 
+### code-review v3.6.1
+
+#### Fixed
+- **Local branch review no longer bases its diff on a stale ref, which folded unrelated commits into the review.** `resolve-scope` hardcoded the diff scope to `main...HEAD`. That `main` is the local branch ref, which every worktree of a clone shares, so a worktree inherits whatever commit the primary checkout last left it on. Because `A...B` diffs from `merge-base(A, B)`, a local ref sitting behind the branch's fork point dragged the merge base backwards and pulled every commit that landed on the base in between into the diff — reviewing other people's work as if it were the branch's. The base is now chosen per run between the local ref and `origin/<base>`: both merge bases are ancestors of HEAD along the base branch, so the ref producing the later one is the true fork point. This is correct under either kind of staleness — a local ref behind the fork point, or an `origin/<base>` behind it because the base has unpushed local commits (where preferring the remote ref would fold those commits in instead) — and under both at once, when the local ref and `origin/<base>` have diverged: whichever tip the branch was cut from yields the deeper merge base, so it is selected regardless of which ref that is. A repo with no remote has no second view and keeps its local ref, so remote-less reviews are unchanged.
+- The base branch is now detected — `origin/HEAD`, then `main`/`master` remotely, then locally — rather than assumed to be `main`, so repositories whose default branch is `master` (or any name `origin/HEAD` reports) resolve their scope correctly instead of failing against a nonexistent `main`. `--base-ref-override` runs through the same selection, and falls back to the local branch when the named base has no remote-tracking ref rather than emitting an `origin/<ref>` that git cannot resolve.
+- `fetch-intent` now reads branch commit subjects from the same fork point instead of the raw local base ref, so intent classification and injection detection no longer receive commits that landed on the base branch and were never part of the change under review.
+
 ### code-review v3.6.0
 
 #### Fixed

plugins/code-review/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "code-review",
   "description": "Code review plugin",
-  "version": "3.6.0",
+  "version": "3.6.1",
   "author": {
     "name": "ClosedLoop",
     "email": "support@closedloop.ai"

plugins/code-review/README.md

@@ -24,12 +24,27 @@ plugins/code-review/
     code-review-worker-graph.md      Graph-aware variant for the cross-file and design reviewers (Impact Analyzer, Bug Hunter B, fast-path, Design Critic); adds read-only codebase-memory-mcp tools — cross-file usage discovery for the cross-file roles, project-structure/dependency-graph analysis (get_architecture, query_graph) for the Design Critic
   commands/
     start.md                         Main /start command (orchestrator)
+    shallow.md                       /shallow wrapper — `/start --depth shallow`
+    deep.md                          /deep wrapper — `/start --depth deep`
+    cost.md                          /cost command — token-cost attribution from session transcripts
+  skills/
+    spawn-reviewers/SKILL.md         Reviewer-fleet spawn/collection contract at stage_20_spawn_reviewers
+    verify-findings/SKILL.md         Finding-verifier fleet dispatch at stage_23_verify_findings (PLN-722)
+    singleton-dispatch/SKILL.md      Single-agent dispatch for stage_11_extract_signals / stage_15_coverage_critic (PLN-725)
+    present-local/SKILL.md           Local-mode presenter at stage_29_present
+    fix/SKILL.md                     Verifies and fixes BLOCKING/HIGH findings from a prior review session
   prompts/
     github-review.md                 GitHub-mode constraints and output steps (loaded conditionally)
+  scripts/
+    dist/cost-report.mjs             Bundled Node cost analyzer for /cost (sources at tools/code-review-cost/)
   tools/
     prompts/shared_prompt.txt        Shared reviewer constraints injected into every agent prompt
     prompts/bha_suffix.txt           Bug Hunter A reviewer persona and focus areas
     prompts/design_critic_suffix.txt Design Critic reviewer role (software-design craftsmanship; always-on at deep tier)
+    prompts/impact_analyzer_prompt.txt   Impact Analyzer reviewer role (FEA-1401 cross-file blast radius; deep tier, signal-gated)
+    prompts/coverage_critic_prompt.txt   Coverage critic role (standard/deep tiers)
+    prompts/signal_extraction_prompt.txt Signal extraction role (standard/deep tiers)
+    prompts/verifier_prompt.txt      Finding-verifier role (falsify-oriented; PLN-722)
     python/code_review_schema.py     Canonical Finding + ResultEnvelope schema + validators (PLN-719)
     python/test_code_review_schema.py  Schema tests + round-trips
     python/code_review_helpers.py    Deterministic helper CLI (parse-diff, hygiene, partition, route, validate, cache, finalize-result, arbitrate-budget, prepare-run, etc.)
@@ -90,34 +105,36 @@ Runs a comprehensive code review. Invokes the full pipeline: diff parsing, hygie
 
 | Argument | Behavior |
 |---|---|
-| _(none)_ | Diff current branch vs `main` |
+| _(none)_ | Review the open PR's diff for the current branch; with no open PR, diff the current branch from its fork point off the default branch |
 | `staged` | Diff only staged (index) changes |
-| `file1 file2 ...` | Diff specific files against `main` |
+| `file1 file2 ...` | Diff specific files from the fork point off the default branch |
 | `123` | Use PR #123's diff (local output, no posting) |
 
+**Base ref resolution.** The default base branch is *detected*, not assumed: the helper reads the `origin/HEAD` symbolic ref, then probes `origin/main` / `origin/master`, then the same names locally, falling back to `main` only when nothing resolves. The diff runs from the fork point rather than a fixed ref — a clone holds both a local `<base>` and an `origin/<base>`, and either can lag the other (a stale local checkout, or unpushed local base commits). The helper takes the merge base of each against `HEAD` and uses whichever ref yields the *later* one, since that is the true fork point. This keeps commits that landed on the base branch after the fork out of the review diff under either kind of staleness.
+
 **Mode flags:**
 
 | Flag | Description |
 |---|---|
 | `--github` | GitHub CI mode: auto-detect PR from branch or accept explicit PR number, post inline comments via file-based handoff |
 | `--github 123` | GitHub CI mode: review PR #123 specifically |
 | `--hygiene-only` | Run only the deterministic hygiene checks. Zero LLM tokens consumed. Fast. |
-| `--base <ref>` | Override the base branch for diffing (default: `main`) |
+| `--base <ref>` | Override the base branch for diffing (default: the repository's detected default branch) |
 | `--since-last-review` | Review only commits added since the last successful review (local mode only) |
 | `--full-review` | Force a full diff even when auto-incremental mode would narrow the scope |
 | `--depth shallow\|standard\|deep` | Reviewer-fleet tier. Default `standard`. See **Depth Tiers** below |
 
 **Examples:**
 
 ```bash
-/start                               # All changes on current branch vs main
+/start                               # Open PR diff, else changes on current branch since its fork point
 /start staged                        # Only staged changes
 /start src/auth.ts src/user.ts       # Specific files
 /start 123                           # PR #123 diff locally
 /start --github                      # CI: auto-detect PR, post comments
 /start --github 123                  # CI: PR #123, post comments
 /start --hygiene-only                # Hygiene checks only
-/start --base develop                # Diff against develop instead of main
+/start --base develop                # Diff against develop instead of the default branch
 /start --since-last-review           # Only new commits since last review
 /start --full-review                 # Disable incremental narrowing
 ```
@@ -309,7 +326,7 @@ Overrides survive across runs while the file content matches and the 90-day TTL
 
 **Re-assert is best-effort against finding_id drift.** Finding IDs are assigned as `<reviewer>_f<index>` where `<index>` is the reviewer's emission position. Across re-runs the LLM may reorder or drop findings, so an override written against `bha_f3` on run N may map to a different finding (or no finding) on run N+1. The content-hash anchor prevents promoting an unrelated finding at a different line — but the common drift case is the override silently no-ops. Two mitigations: (1) re-assert and re-run immediately so the override is honored against the same emission set, and (2) inspect the verify-prepare manifest for `override_hits` / `override_invalidated` to confirm the override landed.
 
-The presenter (local mode `start.md`, GitHub mode `code-review-verifier-stats.md`) surfaces:
+The presenter (local mode: the `present-local` skill, GitHub mode: `github-review.md` Step 6e, which writes `.closedloop-ai/code-review-verifier-stats.md`) surfaces:
 
 - Per-reviewer FP rate (`stats.verification.by_reviewer[*].fp_rate`)
 - Override count per reviewer (`stats.verification.by_reviewer[*].re_asserted`)

plugins/code-review/commands/start.md

@@ -13,7 +13,7 @@ Run a multi-agent code review with partitioned deep review, deterministic hygien
 ## Usage
 
 ```
-/start                              # Review open PR diff for current branch, or main...HEAD if no PR
+/start                              # Review open PR diff for current branch, or the diff since the branch forked from the default branch if no PR
 /start staged                       # Review only staged changes
 /start file1 file2                  # Review specific files
 /start 123                          # Review PR #123 diff locally (no posting)

plugins/code-review/tools/python/code_review_helpers.py

@@ -244,8 +244,8 @@ def _resolve_pr_scope(
     """Resolve diff scope fields for a given PR number.
 
     When *allow_guess_fallback* is ``True`` (explicit ``--pr-number``), a
-    ``CalledProcessError`` from ``gh pr view`` falls back to
-    ``base_ref="main"`` / ``head_ref=current_branch``.  When ``False``
+    ``CalledProcessError`` from ``gh pr view`` falls back to the repo's
+    default branch / ``head_ref=current_branch``.  When ``False``
     (auto-detect path), errors propagate so the caller can revert to branch
     scope.
     """
@@ -256,12 +256,12 @@ def _resolve_pr_scope(
             capture_output=True, text=True, check=True,
         )
         lines = result.stdout.strip().splitlines()
-        base_ref = lines[0].strip() if len(lines) > 0 else "main"
+        base_ref = lines[0].strip() if len(lines) > 0 else _resolve_default_base_ref()
         head_ref = lines[1].strip() if len(lines) > 1 else current_branch
     except subprocess.CalledProcessError:
         if not allow_guess_fallback:
             raise
-        base_ref = "main"
+        base_ref = _resolve_default_base_ref()
         head_ref = current_branch
 
     return {
@@ -295,6 +295,92 @@ def _git_rev_parse(ref: str) -> str | None:
         return None
 
 
+def _resolve_default_base_ref() -> str:
+    """Return the repository's default branch *name* (e.g. ``main``, ``master``).
+
+    Probes, in order: the ``origin/HEAD`` symbolic ref (what the remote
+    reports as its default), then well-known remote branches, then the
+    same names locally for repos with no ``origin``. Falls back to
+    ``main`` when nothing resolves, preserving the historical default.
+
+    Returns a bare branch name, not a ref — callers pair it with
+    :func:`_base_rev` to get the revision to diff against. ``base_ref``
+    travels through ``scope.json`` as a name because consumers such as
+    ``compute-hashes`` origin-qualify it themselves.
+    """
+    try:
+        symbolic = _run_git(
+            ["symbolic-ref", "--short", "refs/remotes/origin/HEAD"],
+        ).strip()
+    except (subprocess.CalledProcessError, FileNotFoundError, OSError):
+        symbolic = ""
+    if symbolic:
+        # "origin/main" -> "main"
+        _, _, name = symbolic.partition("/")
+        if name:
+            return name
+    for candidate in ("main", "master"):
+        if _git_rev_parse(f"origin/{candidate}"):
+            return candidate
+    for candidate in ("main", "master"):
+        if _git_rev_parse(candidate):
+            return candidate
+    return "main"
+
+
+def _merge_base(a: str, b: str) -> str | None:
+    """Return the merge base of *a* and *b*, or ``None`` if there isn't one.
+
+    ``None`` covers both "no common ancestor" and "a ref does not resolve",
+    which callers treat alike: neither yields a usable fork point.
+    """
+    try:
+        return _run_git(["merge-base", a, b]).strip() or None
+    except (subprocess.CalledProcessError, FileNotFoundError, OSError):
+        return None
+
+
+def _is_ancestor(commit: str, descendant: str) -> bool:
+    """Whether *commit* is *descendant* or one of its ancestors."""
+    try:
+        _run_git(["merge-base", "--is-ancestor", commit, descendant])
+        return True
+    except (subprocess.CalledProcessError, FileNotFoundError, OSError):
+        return False
+
+
+def _base_rev(base_ref: str, head_rev: str = "HEAD") -> str:
+    """Return the revision to diff *head_rev* against for branch *base_ref*.
+
+    ``<base>...<head>`` diffs from ``merge-base(<base>, <head>)``, so the base
+    ref matters only through the fork point it produces. A clone holds two
+    views of the same base branch, and either one can lag the other:
+
+    * The local ``<base_ref>`` is shared by every worktree of a clone, so it
+      carries whatever commit the primary checkout last left it on. When it
+      sits behind the fork point, the merge base walks backwards and folds
+      every commit that landed on the base in between into the review diff.
+    * ``origin/<base_ref>`` lags whenever the base has unpushed local commits.
+      Branch off those and the fork point is ahead of the remote ref, which
+      folds the unpushed base commits into the diff instead.
+
+    Both merge bases are ancestors of *head_rev* along the base branch, so the
+    later of the two is the true fork point — take the ref that produces it,
+    which is correct under either kind of staleness. Falls back to whichever
+    ref resolves when only one does (e.g. a remote-less repo, where the local
+    ref is the only truth).
+    """
+    remote_rev = f"origin/{base_ref}"
+    local_mb = _merge_base(base_ref, head_rev)
+    remote_mb = _merge_base(remote_rev, head_rev)
+    if remote_mb is None:
+        return base_ref
+    if local_mb is None:
+        return remote_rev
+    # Equal bases resolve to the remote ref; the two ranges are identical.
+    return remote_rev if _is_ancestor(local_mb, remote_mb) else base_ref
+
+
 # Startup-GC age guard: a PR-head worktree directory is reclaimed as an
 # abort-orphan only once it is older than this. Set far above any real
 # review wall-time (runs are minutes, not hours) so a concurrent in-flight
@@ -4791,7 +4877,7 @@ def cmd_resolve_scope(args: argparse.Namespace) -> int:
         current_branch = "HEAD"
 
     diff_scope = ""
-    base_ref = "main"
+    base_ref = _resolve_default_base_ref()
     head_ref = current_branch
     review_branch = current_branch
     diff_tip = "HEAD"
@@ -4844,21 +4930,23 @@ def cmd_resolve_scope(args: argparse.Namespace) -> int:
                     pr_auto_detected = True
                 except (subprocess.CalledProcessError, FileNotFoundError,
                         OSError, ValueError):
-                    # Any failure: fall back to branch scope
+                    # Any failure: fall back to branch scope. base_ref is
+                    # untouched here — every pr_scope assignment below the
+                    # fetch is unreachable once either raise-point fires.
                     pr_number = None
                     pr_auto_detected = False
-                    diff_scope = "main...HEAD"
+                    diff_scope = f"{_base_rev(base_ref)}...HEAD"
                     scope_kind = "branch"
             else:
-                diff_scope = "main...HEAD"
+                diff_scope = f"{_base_rev(base_ref)}...HEAD"
                 scope_kind = "branch"
         elif scope_args.strip() == "staged":
             diff_scope = "--cached"
             scope_kind = "staged"
         else:
             # Treat scope_args as file paths
             files = scope_args.strip()
-            diff_scope = f"main...HEAD -- {files}"
+            diff_scope = f"{_base_rev(base_ref)}...HEAD -- {files}"
             path_filter = f"-- {files}"
             scope_kind = "file_paths"
 
@@ -4870,11 +4958,12 @@ def cmd_resolve_scope(args: argparse.Namespace) -> int:
     # Apply base-ref override if provided
     if base_ref_override:
         if scope_kind == "pr":
-            diff_scope = f"origin/{base_ref_override}...origin/{head_ref}"
+            override_head = f"origin/{head_ref}"
+            diff_scope = f"{_base_rev(base_ref_override, override_head)}...{override_head}"
         elif path_filter:
-            diff_scope = f"origin/{base_ref_override}...HEAD {path_filter}"
+            diff_scope = f"{_base_rev(base_ref_override)}...HEAD {path_filter}"
         else:
-            diff_scope = f"origin/{base_ref_override}...HEAD"
+            diff_scope = f"{_base_rev(base_ref_override)}...HEAD"
         base_ref = base_ref_override
 
     # Worktree isolation for local PR review. The diff is computed from the
@@ -4987,7 +5076,7 @@ def cmd_fetch_intent(args: argparse.Namespace) -> int:
     elif scope_kind == "branch":
         try:
             result = subprocess.run(
-                ["git", "log", f"{base_ref}..{diff_tip}",
+                ["git", "log", f"{_base_rev(base_ref, diff_tip)}..{diff_tip}",
                  "--oneline", "--no-merges", "--format=%s"],
                 capture_output=True, text=True, check=True,
             )