@caddis/cli 0.2.6 → 0.4.0

package/bundles/antigravity-plugin/scripts/oss_review.py

@@ -0,0 +1,718 @@
+#!/usr/bin/env python3
+"""Cross-vendor code review via any OpenAI-compatible chat-completions endpoint.
+
+A different vendor's model has different blind spots, so a second model reviewing your diff
+catches bugs a same-vendor reviewer misses. This tool is provider-agnostic: point it at
+DeepSeek (default — cheapest + most architecturally distinct from Claude), GLM, OpenRouter,
+or any OpenAI-compatible `/chat/completions` endpoint via three env vars.
+
+Config precedence (highest wins): explicit CLI flag  >  env var  >  provider preset.
+  REVIEW_PROVIDER   preset name: deepseek | glm | openrouter   (default deepseek)
+  REVIEW_BASE_URL   base URL, no trailing /chat/completions    (overrides the preset's URL)
+  REVIEW_MODEL      model id                                   (overrides the preset's model)
+  REVIEW_API_KEY    bearer token                               (REQUIRED — no key ⇒ exit 3)
+
+Future-proofing: model ids and endpoints churn. Two layers protect against that — (1) the PROVIDERS
+table below is the ONE place a renamed model/URL is edited; (2) REVIEW_MODEL / REVIEW_BASE_URL env
+vars always win, so you can point at any new id without touching code. Nothing here is hard-wired.
+
+Usage:
+  python oss_review.py [--range <git range>] [--cwd <repo>] [--provider P] [--base-url U] [--model M]
+    --range   e.g. origin/main..HEAD   (default: the working tree — staged, unstaged AND
+              untracked non-ignored files; a range excludes untracked by design)
+
+Exit codes (fail-closed):
+  0  REVIEW: CLEAN      — no blocking issues
+  1  REVIEW: BLOCKING   — one or more blocking issues
+  2  error              — no diff verdict parsed, git failure, endpoint/parse failure, or the diff
+                          exceeds REVIEW_MAX_DIFF_CHARS (see below — never silently downgraded to CLEAN)
+  3  misconfigured      — REVIEW_API_KEY missing (actionable message on stderr)
+
+Diff-size ceiling: an oversized diff sent to a chat-completions endpoint has been observed, live, to
+come back two different unsafe ways — an empty `content` field on an HTTP 200 (at least fails closed:
+no verdict line ⇒ exit 2), and, worse, a `REVIEW: CLEAN` verdict with no substantive engagement (a
+silent false negative — the model was overwhelmed, not actually reviewing). Rather than risk the second
+case, a diff over REVIEW_MAX_DIFF_CHARS (default 60,000 chars; override via the env var or
+--max-diff-chars) is SPLIT INTO BATCHES on whole-file boundaries and each batch reviewed separately;
+the verdict is CLEAN only if every batch is clean (aggregation is fail-closed). Exit 2 is now the
+narrower case: a SINGLE file larger than the ceiling, which cannot be split, or more batches than
+MAX_REVIEW_BATCHES. Refusal was the original behaviour and is kept for those two - a silent false
+CLEAN is the worst outcome this tool has - but it no longer fires on an ordinary large phase diff.
+
+Stdlib-only (urllib) so it runs anywhere with no pip install.
+"""
+from __future__ import annotations
+
+import argparse
+import http.client
+import json
+import os
+import subprocess
+import sys
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+EXIT_CLEAN = 0
+EXIT_BLOCKING = 1
+EXIT_ERROR = 2
+EXIT_CONFIG = 3
+
+# Provider presets — the SINGLE place a renamed model id or moved endpoint is edited. Adding a new
+# provider (Qwen, a local vLLM, …) is one new row. Callers can always bypass this via env/flags.
+PROVIDERS: dict[str, dict[str, str]] = {
+    "deepseek":   {"base_url": "https://api.deepseek.com",            "model": "deepseek-v4-flash"},
+    "glm":        {"base_url": "https://api.z.ai/api/coding/paas/v4", "model": "glm-5.3"},
+    "openrouter": {"base_url": "https://openrouter.ai/api/v1",        "model": "deepseek/deepseek-v4-flash"},
+}
+DEFAULT_PROVIDER = "deepseek"
+
+# Tried, in order, ONLY when the primary provider fails at transport level (timeout, HTTP
+# error, unparseable response) AND the user did not name a provider explicitly.
+#
+# Why a fallback rather than just picking one vendor: the whole point of this tool is that the
+# reviewer is a DIFFERENT vendor from the author, so collapsing to a single provider means an
+# outage leaves you with no second opinion and — worse — no signal that you have lost one.
+# Observed on serve-sight across three consecutive phases: GLM timed out every time and the
+# sessions recorded "cross-review inconclusive" and moved on. A review that silently does not
+# happen is the same failure class as a diff that silently is not read.
+#
+# Order matters: DeepSeek leads because it has actually found real bugs (a date-boundary bug
+# two same-vendor passes missed, an anchor bug, an orphaned-permission-rows bug). GLM is the
+# spare tyre, not a co-equal — an alternation policy that sends half the phases to a provider
+# that times out buys variety it never actually collects.
+FALLBACK_PROVIDERS: tuple[str, ...] = ("glm",)
+
+# Fail-closed diff-size ceiling (see the module docstring). Observed live: an oversized diff got back
+# an empty `content` field (safe, but wasteful) on one occasion, and — separately, a bulk multi-document
+# diff around 169,000 chars — a `REVIEW: CLEAN` verdict with zero substantive engagement (unsafe: a
+# silent false negative). 60,000 chars is comfortably under both observed failure points while still
+# covering an ordinary multi-file phase diff. REVIEW_MAX_DIFF_CHARS / --max-diff-chars override it.
+DEFAULT_MAX_DIFF_CHARS = 60_000
+
+# Every exception class that means "the request did not come back usable", so main() can route it
+# to provider failover instead of letting it escape.
+#
+# `http.client.HTTPException` is the load-bearing entry and was MISSING. Its subclass
+# `IncompleteRead` — a truncated response body, one of the two live failures behind
+# `.caddis/parking-lot/006` — is NOT a ValueError despite the name suggesting a parse problem; its
+# MRO is (IncompleteRead, HTTPException, Exception). So it escaped the handler entirely, Python
+# exited 1, and 1 is EXIT_BLOCKING. A transport truncation was being reported to scripted callers
+# as "the reviewer found blocking issues" — the same class of bug as the `content: null` crash
+# documented in call_llm, in the same file, unfixed on the sibling path.
+TRANSPORT_ERRORS = (
+    urllib.error.URLError, urllib.error.HTTPError, http.client.HTTPException,
+    RuntimeError, ValueError, TimeoutError, ConnectionError, OSError,
+)
+
+# How many times a failing chunk may be halved and retried before the run gives up. Two levels
+# turns one 8-file batch into at most four 2-file requests, which is enough to clear a
+# size-correlated truncation without turning a broken endpoint into sixteen doomed calls.
+RETRY_SPLIT_DEPTH = 2
+
+
+# A diff over the ceiling is split into batches on whole-FILE boundaries and each batch is
+# reviewed separately, rather than the whole thing being refused. Refusing was correct while
+# there was no alternative — a silent false CLEAN is the worst outcome this tool has — but it
+# put the burden on the author, and the workaround people reached for (`git stash` to split
+# the tree) mutates the working tree mid-verification and cost a hand-resolved merge conflict
+# on serve-sight, 2026-08-03.
+#
+# Batching scales in both directions: a small change is one batch, identical in cost and
+# behaviour to before; a 175k-char phase is three or four, with no judgement call about where
+# to cut. Verdicts aggregate fail-closed (see `aggregate_verdicts`).
+MAX_REVIEW_BATCHES = 12
+
+
+class ConfigError(Exception):
+    """Raised when configuration can't be resolved (unknown provider, or missing API key)."""
+
+
+DEFAULT_KEYS_FILE = "~/.caddis/keys.env"
+
+
+def _parse_keys_file(path: str) -> dict[str, str]:
+    """Parse an INI/``KEY=VALUE`` keys file; missing file yields ``{}``.
+
+    Mirror of claude-harness/scripts/oss_model.py::_parse_keys_file — duplicated (15 lines)
+    rather than imported so this tool stays a single stdlib-only file that runs from any repo
+    the runtime resources are exported into.
+    """
+    p = Path(path).expanduser()
+    if not p.is_file():
+        return {}
+    out: dict[str, str] = {}
+    for raw in p.read_text(encoding="utf-8").splitlines():
+        line = raw.strip()
+        if not line or line.startswith("#") or "=" not in line:
+            continue
+        key, val = line.split("=", 1)
+        out[key.strip()] = val.strip().strip('"').strip("'")
+    return out
+
+
+def resolve_config(
+    args: argparse.Namespace,
+    env: dict[str, str],
+    provider: str | None = None,
+    use_overrides: bool = True,
+) -> tuple[str, str, str]:
+    """(base_url, api_key, model). Precedence: explicit flag > env var > provider preset.
+
+    Never hard-fails on a model rename: a known --provider supplies sane defaults, and
+    REVIEW_MODEL / REVIEW_BASE_URL override them without any code change.
+
+    `provider` names one explicitly (used for fallback). `use_overrides=False` then ignores
+    --model/--base-url/REVIEW_MODEL/REVIEW_BASE_URL, because those were chosen for the PRIMARY
+    provider: pointing DeepSeek's model id at GLM's endpoint would turn a clean fallback into
+    a confusing 404.
+    """
+    provider = (
+        provider or args.provider or env.get("REVIEW_PROVIDER") or DEFAULT_PROVIDER
+    ).strip().lower()
+    preset = PROVIDERS.get(provider, {})
+    if use_overrides:
+        base_url = (args.base_url or env.get("REVIEW_BASE_URL") or preset.get("base_url") or "").rstrip("/")
+        model = args.model or env.get("REVIEW_MODEL") or preset.get("model") or ""
+    else:
+        base_url = (preset.get("base_url") or "").rstrip("/")
+        model = preset.get("model") or ""
+    if not base_url or not model:
+        known = ", ".join(sorted(PROVIDERS))
+        raise ConfigError(
+            f"could not resolve an endpoint for provider {provider!r}. Use a known --provider "
+            f"({known}) or set REVIEW_BASE_URL and REVIEW_MODEL explicitly."
+        )
+    # Key precedence: REVIEW_API_KEY > provider env (DEEPSEEK_API_KEY, …) > OSS_API_KEY >
+    # the caddis keys file — same resolution chain as the claude-oss launcher, so wiring
+    # a provider once (in ~/.caddis/keys.env) lights up both lanes.
+    key_env = f"{provider.upper()}_API_KEY"
+    api_key = ""
+    for name in ("REVIEW_API_KEY", key_env, "OSS_API_KEY"):
+        api_key = (env.get(name) or "").strip()
+        if api_key:
+            break
+    if not api_key:
+        keys_path = env.get("CADDIS_KEYS_FILE") or DEFAULT_KEYS_FILE
+        file_keys = _parse_keys_file(keys_path)
+        for name in (key_env, "OSS_API_KEY"):
+            api_key = (file_keys.get(name) or "").strip()
+            if api_key:
+                break
+    if not api_key:
+        # Name the providers that DO have a key. When DeepSeek is unset but GLM is configured, the
+        # fix is `--provider glm` — and without this line the reader goes hunting for a credential
+        # they already hold, at the least convenient possible moment (a commit pending).
+        keys_path = env.get("CADDIS_KEYS_FILE") or DEFAULT_KEYS_FILE
+        others = [p for p in configured_providers(env) if p != provider]
+        have = (f"\nProviders that DO have a key here: {', '.join(others)}. "
+                f"To use one now: --provider {others[0]}." if others else "")
+        raise ConfigError(
+            f"no API key for provider {provider!r}. Set $REVIEW_API_KEY or ${key_env}, or add\n"
+            f"  {key_env}=<your-key>\n"
+            f"to your keys file ({keys_path}). Override the file path with $CADDIS_KEYS_FILE."
+            + have
+        )
+    return base_url, api_key, model
+
+
+def configured_providers(env: dict[str, str]) -> list[str]:
+    """Every preset provider that currently has a usable key. Never returns or logs a key value.
+
+    Mirror of claude-harness/scripts/oss_model.py::configured_providers — duplicated for the same
+    documented reason as `_parse_keys_file` above: this tool stays a single stdlib-only file that
+    runs from any repo the runtime resources are exported into.
+
+    Exists so a caller can find out BEFORE it needs a key. `/caddis:cross-review` used to discover
+    a missing one mid-task with a commit pending, and a command that fails once at an inconvenient
+    moment does not get retried — so the safety check it provides is quietly lost, and nothing
+    records that it was lost (`.caddis/parking-lot/003`).
+    """
+    keys_path = env.get("CADDIS_KEYS_FILE") or DEFAULT_KEYS_FILE
+    file_keys = _parse_keys_file(keys_path)
+    generic = ((env.get("REVIEW_API_KEY") or env.get("OSS_API_KEY") or "").strip()
+               or (file_keys.get("OSS_API_KEY") or "").strip())
+    out = []
+    for name in sorted(PROVIDERS):
+        key_env = f"{name.upper()}_API_KEY"
+        if generic or (env.get(key_env) or "").strip() or (file_keys.get(key_env) or "").strip():
+            out.append(name)
+    return out
+
+
+def _untracked_files(cwd: str) -> list[str]:
+    """New, non-ignored files git is not tracking yet. Empty list on any git failure."""
+    out = subprocess.run(
+        ["git", "ls-files", "--others", "--exclude-standard", "-z"],
+        cwd=cwd, capture_output=True, text=True,
+        encoding="utf-8", errors="replace", timeout=30,
+    )
+    if out.returncode != 0:
+        return []
+    return [p for p in out.stdout.split("\0") if p]
+
+
+def _untracked_diff(path: str, cwd: str) -> str:
+    """A synthetic add-hunk for one untracked file, WITHOUT touching the index.
+
+    `git add -N` would also work and is shorter, but it WRITES to the index — unacceptable
+    in a review tool that routinely runs mid-session against a tree the user is still
+    working in. `--no-index` gets the same diff with no side effect.
+
+    Note the exit code: `git diff --no-index` returns 1 when the files differ, which is the
+    normal outcome here. Only >1 is a real failure.
+    """
+    null = "/dev/null" if os.name != "nt" else "NUL"
+    out = subprocess.run(
+        ["git", "diff", "--no-index", "--", null, path],
+        cwd=cwd, capture_output=True, text=True,
+        encoding="utf-8", errors="replace", timeout=30,
+    )
+    return out.stdout if out.returncode <= 1 else ""
+
+
+def get_diff(rng: str | None, cwd: str, include_untracked: bool = True) -> str:
+    """The unified diff for `rng` (e.g. 'origin/main..HEAD'), or the working tree when None.
+
+    Working tree = staged + unstaged against HEAD, PLUS untracked non-ignored files.
+
+    Untracked files are included because `git diff HEAD` reports only paths git already
+    tracks, so a change consisting entirely of NEW files produced an empty diff — and the
+    empty-diff branch in main() prints `REVIEW: CLEAN` and exits 0 without ever calling the
+    model. A brand-new module is exactly the kind of change most worth reviewing, and it was
+    the kind that silently got no review at all. Found on serve-sight 2026-08-02: four
+    consecutive cross-review runs over a phase of mostly-new files reported CLEAN, and the
+    real findings only appeared once the author happened to run `git add -A` first.
+
+    Untracked files are NOT included for an explicit `rng` — a commit range is a span of
+    history, and files that were never committed are correctly outside it.
+    """
+    cmd = ["git", "diff", rng] if rng else ["git", "diff", "HEAD"]
+    # encoding pinned to UTF-8: git emits UTF-8, but text=True would decode with the locale
+    # codepage (cp1252 on Windows) — a non-ASCII diff then kills the reader thread and
+    # subprocess hands back stdout=None with returncode 0.
+    out = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True,
+                         encoding="utf-8", errors="replace", timeout=30)
+    if out.returncode != 0:
+        raise RuntimeError(f"git diff failed: {out.stderr.strip()}")
+    if rng or not include_untracked:
+        return out.stdout
+    extra = "".join(_untracked_diff(p, cwd) for p in _untracked_files(cwd))
+    # Concatenated BEFORE returning so the caller's REVIEW_MAX_DIFF_CHARS ceiling measures
+    # the true payload rather than the tracked-only fraction of it.
+    return out.stdout + extra
+
+
+def split_diff_by_file(diff_text: str) -> list[str]:
+    """The diff cut into per-file chunks on `diff --git` boundaries.
+
+    Splitting on file boundaries, never mid-file: half a file's hunks is worse than no
+    review of it, because the model confidently reasons about code it cannot see.
+    """
+    if not diff_text.strip():
+        return []
+    parts: list[str] = []
+    current: list[str] = []
+    for line in diff_text.splitlines(keepends=True):
+        if line.startswith("diff --git ") and current:
+            parts.append("".join(current))
+            current = [line]
+        else:
+            current.append(line)
+    if current:
+        parts.append("".join(current))
+    return parts
+
+
+def batch_diff(diff_text: str, max_chars: int) -> list[str]:
+    """Pack whole files into as few batches as possible, each under `max_chars`.
+
+    A single file larger than the ceiling gets its own batch and is sent anyway — it is
+    over the limit either way, and reviewing it is strictly better than skipping it. The
+    caller reports how many batches were used so an oversized single file is visible.
+    """
+    files = split_diff_by_file(diff_text)
+    if not files:
+        return []
+    batches: list[str] = []
+    current = ""
+    for chunk in files:
+        if current and len(current) + len(chunk) > max_chars:
+            batches.append(current)
+            current = chunk
+        else:
+            current += chunk
+    if current:
+        batches.append(current)
+    return batches
+
+
+def aggregate_verdicts(verdicts: list[bool | None]) -> bool | None:
+    """Fail-closed roll-up: any BLOCKING wins; CLEAN only if EVERY batch was clean.
+
+    A batch with no parseable verdict (`None`) poisons the result to `None`, which the
+    caller maps to EXIT_ERROR — the same fail-closed rule as the single-batch path. An
+    unreviewed batch must never be able to produce a CLEAN overall.
+    """
+    if not verdicts:
+        return None
+    if any(v is False for v in verdicts):
+        return False
+    if any(v is None for v in verdicts):
+        return None
+    return True
+
+
+def current_branch(cwd: str) -> str:
+    try:
+        out = subprocess.run(
+            ["git", "rev-parse", "--abbrev-ref", "HEAD"],
+            cwd=cwd, capture_output=True, text=True,
+            encoding="utf-8", errors="replace", timeout=10,
+        )
+        return out.stdout.strip() or "the current branch"
+    except Exception:
+        return "the current branch"
+
+
+def build_review_prompt(diff_text: str, branch: str, rng: str | None) -> str:
+    """Adversarial review prompt — ported from docket runner._review_prompt (kept in sync).
+
+    Self-contained: the criteria are inline so a real review always happens, and it ends with
+    a single machine-parseable verdict line the caller maps to an exit code.
+    """
+    scope = rng or "the working tree (staged, unstaged and new untracked files)"
+    return (
+        f"Perform an adversarial code review of the changes on branch '{branch}' ({scope}). "
+        "The unified diff is provided below. Judge, in priority order: (1) correctness — logic "
+        "bugs, wrong results, missed edge cases; (2) tests — would a test fail without this "
+        "change, and are the stated behaviors covered; (3) security — injection, auth, secret "
+        "exposure, unvalidated input; (4) conventions — the repo's stated rules; (5) simplicity. "
+        "Classify each issue as blocking (must fix before merge), should-fix, or nit. Be "
+        "specific: cite the file and line. Do not invent issues; if the diff is clean, say so.\n\n"
+        "End with EXACTLY one line and nothing after it: `REVIEW: CLEAN` (no blocking issues) or "
+        "`REVIEW: BLOCKING` (one or more blocking issues).\n\n"
+        "----- BEGIN DIFF -----\n"
+        f"{diff_text}\n"
+        "----- END DIFF -----"
+    )
+
+
+def classify_verdict(text: str) -> bool | None:
+    """CLEAN→True, BLOCKING→False, neither→None. Fail-closed: any blocking signal wins."""
+    u = text.upper()
+    if "REVIEW: BLOCKING" in u:
+        return False
+    if "REVIEW: CLEAN" in u:
+        return True
+    return None
+
+
+def call_llm(base_url: str, api_key: str, model: str, prompt: str, timeout: int = 180) -> str:
+    """POST to {base_url}/chat/completions (OpenAI-compatible) → the assistant message text.
+
+    Raises on transport, HTTP, or response-shape failure so main() maps it to EXIT_ERROR.
+    """
+    payload = json.dumps({
+        "model": model,
+        "messages": [{"role": "user", "content": prompt}],
+        "temperature": 0,
+        "stream": False,
+    }).encode("utf-8")
+    req = urllib.request.Request(
+        f"{base_url}/chat/completions",
+        data=payload,
+        headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+        method="POST",
+    )
+    with urllib.request.urlopen(req, timeout=timeout) as resp:
+        data = json.loads(resp.read().decode("utf-8"))
+    try:
+        content = data["choices"][0]["message"]["content"]
+    except (KeyError, IndexError, TypeError) as exc:
+        raise RuntimeError(f"unexpected response shape from endpoint: {exc}") from exc
+    # An HTTP 200 carrying empty or null content is a REAL observed failure (see the module
+    # docstring: an oversized diff came back with an empty `content` field). Returning it
+    # unraised had two bad consequences, both silent:
+    #   - it never reached the `except` in main(), so the provider FAILOVER never fired — the
+    #     run died having never tried the spare, which is the exact case failover exists for;
+    #   - `content: null` (legal in OpenAI-compatible responses) returned None, and
+    #     `classify_verdict(None)` then raised AttributeError. Uncaught, Python exits 1 —
+    #     which is EXIT_BLOCKING. A crash was being reported to scripted callers as "the
+    #     reviewer found blocking issues."
+    # Raising here routes both into the existing retry/failover path and kills the crash.
+    if not isinstance(content, str) or not content.strip():
+        raise RuntimeError(
+            "endpoint returned HTTP 200 with empty or absent message content "
+            f"(type={type(content).__name__}) — treating as a provider failure"
+        )
+    return content
+
+
+def _force_utf8_stdio() -> None:
+    """Make stdout/stderr survive a review written by a model, on any console.
+
+    A review is prose, and prose from a model contains →, —, ’, … or ≥ sooner or later. A
+    Windows console defaults to cp1252, where `print(review)` raises UnicodeEncodeError —
+    AFTER the provider round-trip has already been paid for (~8 minutes, observed) and, because
+    the print sits inside the per-batch loop, killing every remaining batch with it. Worse, the
+    caller sees a traceback and a non-zero exit, which is indistinguishable from a provider
+    timeout: the documented response is "switch vendor and retry", so the next move burns
+    another full review on the other provider and hits the identical crash, because the cause is
+    local stdout encoding, not the provider.
+
+    `errors="replace"` is the load-bearing part: a mangled arrow is a far better outcome than a
+    discarded review. Guarded because stdout may already have been swapped for something without
+    `.reconfigure` (pytest's capture, a pipe wrapper) — in which case there is nothing to fix.
+    """
+    for stream in (sys.stdout, sys.stderr):
+        reconfigure = getattr(stream, "reconfigure", None)
+        if reconfigure is None:
+            continue
+        try:
+            reconfigure(encoding="utf-8", errors="replace")
+        except Exception:  # pragma: no cover — defensive; never lose a review to this
+            pass
+
+
+def main(argv: list[str] | None = None, env: dict[str, str] | None = None) -> int:
+    _force_utf8_stdio()
+    env = os.environ if env is None else env
+    parser = argparse.ArgumentParser(description="Cross-vendor code review via an OpenAI-compatible endpoint.")
+    parser.add_argument("--range", dest="range", default=None,
+                        help="git range to review (e.g. origin/main..HEAD); default = working tree")
+    parser.add_argument("--cwd", default=".", help="repo directory (default: cwd)")
+    parser.add_argument("--provider", default=None,
+                        help=f"preset: {', '.join(sorted(PROVIDERS))} (default {DEFAULT_PROVIDER}); env REVIEW_PROVIDER")
+    parser.add_argument("--base-url", dest="base_url", default=None, help="override REVIEW_BASE_URL / the preset")
+    parser.add_argument("--model", default=None, help="override REVIEW_MODEL / the preset")
+    parser.add_argument("--max-diff-chars", type=int, default=None,
+                        help=f"override REVIEW_MAX_DIFF_CHARS (default {DEFAULT_MAX_DIFF_CHARS})")
+    parser.add_argument("--check-config", action="store_true",
+                        help="report whether a usable key exists and exit — no diff, no LLM call. "
+                             "Exit 0 = ready, 3 = nothing configured.")
+    args = parser.parse_args(argv)
+
+    # --check-config is deliberately NON-INTERACTIVE. This script runs headless — from CI, from a
+    # docket runner, from a `claude -p` session — where a prompt on stdin does not ask a question,
+    # it hangs forever. The ASKING belongs in `/caddis:cross-review`, which is driven by an agent
+    # that can actually talk to a human. This half just answers "is it configured?" cheaply enough
+    # to call before you need it, instead of finding out mid-commit.
+    if args.check_config:
+        available = configured_providers(env)
+        if not available:
+            sys.stderr.write(
+                "no second-vendor API key is configured anywhere.\n"
+                "  Cross-review needs one — the whole point is a reviewer that does not share this\n"
+                "  model's blind spots, so a same-vendor fallback would be no review at all.\n"
+                f"  Add a key to {env.get('CADDIS_KEYS_FILE') or DEFAULT_KEYS_FILE}, e.g.\n"
+                "    DEEPSEEK_API_KEY=<your-key>\n")
+            return EXIT_CONFIG
+        print(f"cross-review is ready. Providers with a key: {', '.join(available)}")
+        return EXIT_CLEAN
+
+    try:
+        base_url, api_key, model = resolve_config(args, env)
+        primary_provider = (
+            args.provider or env.get("REVIEW_PROVIDER") or DEFAULT_PROVIDER
+        ).strip().lower()
+    except ConfigError as exc:
+        sys.stderr.write(f"{exc}\n")
+        return EXIT_CONFIG
+
+    try:
+        diff_text = get_diff(args.range, args.cwd)
+    except Exception as exc:
+        sys.stderr.write(f"could not read diff: {exc}\n")
+        return EXIT_ERROR
+
+    if not diff_text.strip():
+        # Says what was scanned, not just that nothing was found. This branch used to fire
+        # on a tree full of NEW files (they are invisible to `git diff HEAD`) and report
+        # CLEAN without calling the model — a silent false negative that reads identically
+        # to a real pass. Untracked files are in scope now; naming the scope keeps the
+        # message honest if that ever regresses.
+        print(f"No changes to review in {args.range or 'the working tree (incl. untracked)'}.")
+        print("REVIEW: CLEAN")
+        return EXIT_CLEAN
+
+    max_diff_chars = args.max_diff_chars
+    if max_diff_chars is None:
+        env_val = (env.get("REVIEW_MAX_DIFF_CHARS") or "").strip()
+        max_diff_chars = int(env_val) if env_val.isdigit() else DEFAULT_MAX_DIFF_CHARS
+    batches = batch_diff(diff_text, max_diff_chars) or [diff_text]
+    # Batching only helps when the diff HAS file boundaries to cut on. A single file (or a
+    # blob with no `diff --git` headers) larger than the ceiling cannot be split, and sending
+    # it anyway would reintroduce exactly the silent false-CLEAN the ceiling exists to stop.
+    # Refuse, as before — the guarantee is absolute: no prompt over the ceiling is ever sent.
+    if any(len(b) > max_diff_chars for b in batches):
+        biggest = max(batches, key=len)
+        first_line = biggest.splitlines()[0][:120] if biggest.strip() else "(no file header)"
+        sys.stderr.write(
+            f"diff is {len(diff_text):,} chars and contains a single chunk of {len(biggest):,}, "
+            f"over the {max_diff_chars:,}-char review ceiling — refusing rather than risk an "
+            "unreviewable prompt silently coming back CLEAN.\n"
+            f"  Largest chunk starts: {first_line}\n"
+            + (
+                # The working tree includes untracked files by design (that fix stopped a
+                # false-CLEAN on all-new-file phases), which means an agent's own uncommitted
+                # plans/PRDs are counted as "the diff". Observed live: 77k of untracked
+                # .caddis/ markdown pushed a real phase over the ceiling. Reviewing the phase's
+                # commit range is both the workaround AND the more accurate scope, so name it
+                # as a literal command instead of leaving each caller to rediscover it.
+                "  You are reviewing the WORKING TREE, which includes untracked files — "
+                "uncommitted docs/plans count toward this total.\n"
+                "  If the phase is already committed, review its commits instead (usually the "
+                "scope you actually want):\n"
+                "      --range HEAD~1..HEAD\n"
+                if args.range is None
+                else "  Narrow --range to a smaller commit span, or review file-by-file.\n"
+            )
+            + "  Override (not recommended without knowing your endpoint's real limit): "
+            "--max-diff-chars <n> or $REVIEW_MAX_DIFF_CHARS.\n"
+        )
+        return EXIT_ERROR
+    if len(batches) > MAX_REVIEW_BATCHES:
+        # Loud, never a silent truncation: a capped review that reads like a complete one is
+        # the same class of lie as an oversized prompt coming back CLEAN.
+        sys.stderr.write(
+            f"diff is {len(diff_text):,} chars — {len(batches)} batches, over the "
+            f"{MAX_REVIEW_BATCHES}-batch cap. Refusing rather than reviewing part of it and "
+            "reporting a verdict that looks whole.\n"
+            "  Narrow --range to a smaller commit span, or raise --max-diff-chars if you know "
+            "your endpoint's real limit.\n"
+        )
+        return EXIT_ERROR
+    if len(batches) > 1:
+        print(
+            f"[cross-review] diff is {len(diff_text):,} chars, over the {max_diff_chars:,} "
+            f"ceiling — reviewing in {len(batches)} batches split on file boundaries. "
+            "Verdict is CLEAN only if every batch is clean."
+        )
+
+    branch = current_branch(args.cwd)
+
+    # Fall back to another vendor only when the user did NOT name one. If they asked for a
+    # specific provider, silently reviewing with a different one would misreport who checked
+    # the code — the single fact this tool exists to establish.
+    # "Explicit" must include a custom ENDPOINT, not just a named preset. Someone who set
+    # REVIEW_BASE_URL/REVIEW_MODEL (the documented future-proofing path) has named exactly who
+    # should review their code; falling back to the GLM preset would misreport the reviewer —
+    # the one fact this tool exists to establish.
+    explicit = bool(
+        args.provider or env.get("REVIEW_PROVIDER")
+        or args.base_url or env.get("REVIEW_BASE_URL")
+        or args.model or env.get("REVIEW_MODEL")
+    )
+    chain: list[tuple[str, str, str, str]] = [(primary_provider, base_url, api_key, model)]
+    if not explicit:
+        for name in FALLBACK_PROVIDERS:
+            if name == primary_provider:
+                continue
+            try:
+                fb_url, fb_key, fb_model = resolve_config(args, env, provider=name, use_overrides=False)
+            except ConfigError:
+                continue  # no key for the spare — not an error, just nothing to fall back to
+            chain.append((name, fb_url, fb_key, fb_model))
+
+    def _try_chain(prompt: str, label: str) -> str | None:
+        """One payload against every provider in the chain. None when all of them failed."""
+        for idx, (name, url, key, mdl) in enumerate(chain):
+            try:
+                review = call_llm(url, key, mdl, prompt)
+            except TRANSPORT_ERRORS as exc:
+                remaining = len(chain) - idx - 1
+                sys.stderr.write(
+                    f"review request failed{label} ({mdl} @ {url}): {exc}\n"
+                    + (
+                        f"  falling back to {chain[idx + 1][0]}...\n"
+                        if remaining
+                        else "  If the model id was renamed, set REVIEW_MODEL to the current id "
+                             "(env overrides the preset).\n"
+                    )
+                )
+                continue
+            if idx:
+                # Loud on purpose. A verdict from a different vendor than the one requested is
+                # a material fact about the review, not an implementation detail — it belongs
+                # in the transcript the phase report quotes.
+                print(f"[cross-review] {chain[0][0]} unavailable; reviewed by {name} ({mdl}).")
+            return review
+        return None
+
+    def _review_chunk(chunk: str, label: str, depth: int = 0) -> list[str] | None:
+        """Review `chunk`, halving it and retrying once per level if every provider failed.
+
+        WHY HALVE RATHER THAN JUST FAIL. Both observed failures were transport truncation on a
+        long reply — an `http.client.IncompleteRead` on one provider and an empty HTTP 200 body on
+        another — and both were SIZE-CORRELATED. Retrying the same payload against the same
+        endpoint reproduces the same overflow; retrying a smaller one usually does not. The whole
+        run used to abort here, which is safe but throws away a review the user has already
+        waited minutes for.
+
+        Bounded at RETRY_SPLIT_DEPTH levels and never splits a single file, so it cannot loop and
+        cannot silently shrink a diff to nothing. Returns None when a chunk that cannot be split
+        any further still fails — the abort path stays reachable.
+        """
+        review = _try_chain(build_review_prompt(chunk, branch, args.range), label)
+        if review is not None:
+            return [review]
+        files = split_diff_by_file(chunk)
+        if depth >= RETRY_SPLIT_DEPTH or len(files) < 2:
+            return None
+        half = len(files) // 2
+        sys.stderr.write(
+            f"[cross-review] retrying{label} as 2 smaller chunks ({half} + {len(files) - half} "
+            "file(s)) — both observed failures were size-correlated truncation.\n")
+        out: list[str] = []
+        for part_no, part in enumerate(("".join(files[:half]), "".join(files[half:])), 1):
+            got = _review_chunk(part, f"{label} part {part_no}", depth + 1)
+            if got is None:
+                return None
+            out += got
+        return out
+
+    verdicts: list[bool | None] = []
+    for batch_no, batch in enumerate(batches, 1):
+        label = f" (batch {batch_no}/{len(batches)})" if len(batches) > 1 else ""
+        if len(batches) > 1:
+            files = len(split_diff_by_file(batch))
+            print(f"\n===== batch {batch_no}/{len(batches)} — {files} file(s), "
+                  f"{len(batch):,} chars =====")
+
+        reviews = _review_chunk(batch, label)
+        if reviews is None:
+            # Every provider failed, and splitting did not rescue it. Do NOT continue to the next
+            # batch: a partial sweep that still prints a verdict is the failure mode this whole
+            # tool guards against.
+            sys.stderr.write(f"batch {batch_no}/{len(batches)} could not be reviewed — aborting.\n")
+            return EXIT_ERROR
+
+        for review in reviews:
+            print(review)
+            verdicts.append(classify_verdict(review))
+
+    verdict = aggregate_verdicts(verdicts)
+    if len(batches) > 1:
+        # A LEDGER, not a summary. The live report behind `.caddis/parking-lot/006` said the
+        # reader's only defence was "count the REVIEW: lines against the announced batch count" —
+        # so print that arithmetic instead of making a human do it. `reviewed` can exceed
+        # `announced` when a batch was split and retried, and that is worth seeing.
+        clean = sum(1 for v in verdicts if v is True)
+        blocking = sum(1 for v in verdicts if v is False)
+        unparsed = sum(1 for v in verdicts if v is None)
+        print(f"\n[cross-review] announced {len(batches)} batches · reviewed {len(verdicts)} · "
+              f"clean {clean} · blocking {blocking} · no parseable verdict {unparsed}")
+        if unparsed:
+            print("[cross-review] a chunk with no verdict is NOT a pass — the overall result "
+                  "below is fail-closed.")
+    if verdict is True:
+        return EXIT_CLEAN
+    if verdict is False:
+        return EXIT_BLOCKING
+    sys.stderr.write("no REVIEW: CLEAN|BLOCKING verdict line found in the model output\n")
+    return EXIT_ERROR
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())