alissa-tools-github-revloop 0.16.2__tar.gz → 0.16.4__tar.gz

{alissa_tools_github_revloop-0.16.2/src/main/alissa_tools_github_revloop.egg-info → alissa_tools_github_revloop-0.16.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: alissa-tools-github-revloop
-Version: 0.16.2
+Version: 0.16.4
 Summary: ALISSA-TOOLS-GITHUB-REVLOOP
 Home-page: https://alissa.app
 Author: Fahera

{alissa_tools_github_revloop-0.16.2 → alissa_tools_github_revloop-0.16.4}/src/main/alissa/tools/github/revloop/__main__.py

@@ -94,6 +94,14 @@ def build_parser() -> argparse.ArgumentParser:
     over.add_argument("--hub-template", metavar="TEMPLATE")
     over.add_argument("--agent-profile", metavar="NAME")
     over.add_argument("--reviewer-login", metavar="LOGIN")
+    over.add_argument(
+        "--reviewer-token-env",
+        metavar="VAR",
+        help="NAME of the environment variable holding the reviewer "
+        "identity's GitHub token (never the token itself). Set it and every "
+        "`gh` call runs under that credential explicitly instead of whatever "
+        "the container happened to export",
+    )
     over.add_argument("--state-path", type=Path, metavar="PATH")
     over.add_argument(
         "--on-missing-review-task",
@@ -143,6 +151,7 @@ def overrides_from(args: argparse.Namespace) -> dict:
         "hub_template": args.hub_template,
         "agent_profile": args.agent_profile,
         "reviewer_login": args.reviewer_login,
+        "reviewer_token_env": args.reviewer_token_env,
         "state_path": args.state_path,
         "on_missing_review_task": args.on_missing_review_task,
         "on_missing_hub": args.on_missing_hub,

{alissa_tools_github_revloop-0.16.2 → alissa_tools_github_revloop-0.16.4}/src/main/alissa/tools/github/revloop/config.py

@@ -15,10 +15,89 @@ daemons over different workspaces on the same machine, each pointed with
 from __future__ import annotations
 
 import json
+import re
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Mapping
 
+# A POSIX-ish environment variable name -- what `reviewer_token_env` must be.
+_ENV_NAME_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
+
+# The per-character half of the same rule, used to name the offending
+# characters in an error without echoing the value (see _describe).
+_ENV_CHAR_RE = re.compile(r"[A-Za-z0-9_]")
+
+# ...which, on its own, accepts every GitHub credential format: `ghp_`, `gho_`,
+# `ghu_`, `ghs_`, `ghr_` and `github_pat_` tokens are `[A-Za-z0-9_]` throughout
+# and therefore valid identifiers. So the shape has to be rejected explicitly,
+# or the one mistake this validation exists to catch -- pasting the token where
+# its variable's NAME belongs -- sails through.
+_TOKEN_SHAPE_RE = re.compile(r"^(gh[pousr]_|github_pat_)", re.IGNORECASE)
+
+# Generous on purpose. A tighter ceiling (32 was tried) rejects names an
+# operator can plausibly write -- `REVLOOP_REVIEWER_GITHUB_TOKEN_ENV` is 33 --
+# and this bound is NOT what catches credentials: MIN_SECRET_RUN_LENGTH is,
+# and it catches a 40-character hex token on its shape whatever the ceiling
+# says. So the ceiling exists only to refuse the obviously absurd, and it is
+# deliberately kept OUT of `_is_credential_shaped` (see there).
+MAX_ENV_NAME_LENGTH = 40
+
+# Length is not the whole discriminator, because a secret can be short enough
+# to fit under the ceiling. The shape is: a long, undifferentiated run of
+# alphanumerics. Names are conventionally SCREAMING_SNAKE and separated -- the
+# underscore is what a legacy 40-hex token and a generic alnum API key lack.
+MIN_SECRET_RUN_LENGTH = 20
+
+
+def _is_credential_shaped(value: str) -> bool:
+    """Whether a `reviewer_token_env` value looks like a pasted secret.
+
+    A heuristic, and deliberately one that errs toward accusing: the cost of a
+    false positive is renaming a variable, and the cost of a false negative is
+    a live credential sitting in a config file on a shared volume.
+
+    Length is deliberately NOT one of the clauses. It was, and it made a merely
+    LONG name get classified as a credential -- rejected, redacted, and told to
+    rotate itself. Over-length is a separate refusal with its own message; only
+    an actual credential SHAPE drives redaction.
+    """
+    return bool(
+        _TOKEN_SHAPE_RE.match(value)
+        or (
+            len(value) >= MIN_SECRET_RUN_LENGTH
+            and "_" not in value
+            and value.isalnum()
+        )
+    )
+
+
+def _describe(value: str) -> str:
+    """Say what is wrong with a rejected value without reprinting a secret.
+
+    `__main__` prints this to stderr as `config error: …`, straight into the
+    container log, so a pasted credential must never be echoed there. But the
+    other way to fail this check is an ordinary typo
+    (`REVLOOP-REVIEWER-GH-TOKEN`), and there the value IS the diagnostic.
+
+    Redacting everything loses the typo case; echoing whatever the GitHub-shape
+    heuristic does not recognise hands back the container-log disclosure for
+    every non-GitHub secret (`sk-proj-…`, `xoxb-…`, `glpat-…` all carry
+    punctuation and match no clause above). So neither branch prints the value:
+    what an operator needs is WHICH CHARACTERS were rejected, and those can be
+    named with their positions and nothing else.
+    """
+    if _is_credential_shaped(value):
+        return f"a {len(value)}-character value that looks like a credential"
+    offenders = sorted({c for c in value if not _ENV_CHAR_RE.match(c)})
+    if offenders:
+        where = ", ".join(
+            f"{c!r} at {', '.join(str(i) for i, ch in enumerate(value) if ch == c)}"
+            for c in offenders
+        )
+        return f"a {len(value)}-character value containing {where}"
+    return f"a {len(value)}-character value"
+
+
 # What to do when a PR has a pending review request but no matching Alissa
 # review task (CR2 is implementer-side, so a third-party PR may not have one).
 ON_MISSING_SPAWN = "spawn_anyway"  # review anyway, PR URL carries the context
@@ -44,6 +123,7 @@ CONFIG_KEYS = (
     "operators",
     "agent_profile",
     "reviewer_login",
+    "reviewer_token_env",
     "state_path",
     "on_missing_review_task",
     "on_missing_hub",
@@ -117,6 +197,15 @@ class Config:
     agent_profile: str = "claude"
     reviewer_login: str | None = None  # None -> resolve once via `gh api user`
 
+    # NAME of the environment variable holding the reviewer identity's GitHub
+    # token -- never the token itself, which has no business in a config file
+    # on a mounted volume. Set it and every `gh` call the daemon makes runs
+    # under that credential explicitly, with the container's inherited
+    # GH_TOKEN/GITHUB_TOKEN stripped; leave it None and the daemon inherits,
+    # which is the pre-#51 behaviour and warns loudly at preflight because a
+    # shared container is exactly where inheritance picks the wrong identity.
+    reviewer_token_env: str | None = None
+
     # None means "derive from the workspace" -- read `state_db` for the
     # resolved location, never this field.
     state_path: Path | None = None
@@ -247,6 +336,33 @@ class Config:
             # state of a working loop -- an alarm that always fires is noise.
             raise ValueError(f"reap_session_cap must be >= 1, got {session_cap}")
 
+        token_env = raw.get("reviewer_token_env")
+        if token_env is not None:
+            token_env = str(token_env).strip()
+            # The overwhelmingly likely way to get here is pasting the TOKEN in
+            # place of the variable's name, which would leave a live credential
+            # in a config file on a shared volume and then fail as "unset
+            # variable" — an error that reads as though the secret were missing
+            # rather than exposed. So the check is two-sided: it must LOOK like
+            # a name, and it must not look like a credential. The value itself
+            # is never echoed back (see _redact).
+            if (
+                not _ENV_NAME_RE.match(token_env)
+                or _is_credential_shaped(token_env)
+                or len(token_env) > MAX_ENV_NAME_LENGTH
+            ):
+                rotate = (
+                    " If that is a credential, rotate it: it is now in a config file."
+                    if _is_credential_shaped(token_env)
+                    else ""
+                )
+                raise ValueError(
+                    f"reviewer_token_env must be an environment variable NAME "
+                    f"(e.g. 'REVLOOP_REVIEWER_GH_TOKEN') of at most "
+                    f"{MAX_ENV_NAME_LENGTH} characters, not a value or a "
+                    f"token — got {_describe(token_env)}.{rotate}"
+                )
+
         state_path = raw.get("state_path")
         return cls(
             workspace_root=Path(workspace_root),
@@ -257,6 +373,7 @@ class Config:
             operators=operators,
             agent_profile=raw.get("agent_profile", "claude"),
             reviewer_login=raw.get("reviewer_login"),
+            reviewer_token_env=token_env or None,
             state_path=Path(state_path).expanduser() if state_path else None,
             on_missing_review_task=mode,
             on_missing_hub=hub_mode,