@salesforce/afv-skills 1.14.0 → 1.16.0

package/skills/investigating-agentforce-architecture/scripts/rest_client.py

@@ -0,0 +1,763 @@
+"""Tooling + Data REST client scaffolding with security-critical primitives.
+
+This module ships the HTTP primitives used by every REST-path caller in the
+skill. Two security invariants are enforced here and must not be bypassed:
+
+the Authorization header is STRIPPED from any cross-host redirect.
+Python's default HTTPRedirectHandler blindly forwards all request headers
+(including Authorization) to the redirect target. A compromised or
+attacker-controlled edge that returns a 302 to an arbitrary host would
+otherwise receive the bearer token. We subclass HTTPRedirectHandler to
+strip Authorization whenever the redirect target hostname differs from
+the original. Callers MUST use `build_opener()` — never `urllib.request.
+urlopen` directly, which wires the default redirect handler.
+
+access tokens MUST NEVER appear in exception strings, tracebacks,
+or logged output. `redact_error(exc)` returns a safe string representation
+with `Authorization: Bearer ...` scrubbed. Every `except` that surfaces
+HTTP / subprocess error text runs the text through `redact_error` first.
+Never call `log.exception()` on an object that carries request headers.
+"""
+from __future__ import annotations
+
+import email.utils
+import functools
+import json
+import logging
+import re
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+from typing import Any, Callable, Tuple
+
+
+# stdlib logger for transient-HTTP backoff breadcrumbs. The skill has
+# no global logging config — we emit at DEBUG and leave handler/level wiring
+# to callers. Absent any config, Python's "last-resort" handler suppresses
+# DEBUG records, so this is a no-op in production until someone enables it.
+logger = logging.getLogger(__name__)
+
+
+class RestClientError(RuntimeError):
+    """REST request failed — always constructed with a redacted message."""
+
+
+# -----------------------------------------------------------------------------
+# error redaction
+# -----------------------------------------------------------------------------
+# Three patterns cover the bulk of token-leakage surfaces:
+# 1. `Authorization: Bearer <token>` in any string context (header echo,
+# exception repr, stringified HTTP error).
+# 2. `accessToken=<token>` / `access_token=<token>` in URL-encoded bodies
+# or query strings (`sf org display` errors sometimes echo these).
+# 3. `"accessToken":"<token>"` or `"access_token":"<token>"` in JSON
+# payload echoes.
+#
+# Regexes are intentionally permissive on the token character class
+# (anything non-whitespace, non-quote, non-ampersand) — we'd rather
+# over-redact than miss a token with an unexpected encoding.
+
+_AUTH_HEADER_RE = re.compile(
+    r"(Authorization\s*:\s*Bearer\s+)\S+",
+    flags=re.IGNORECASE,
+)
+# Matches access[_]?Token=<value> in url-encoded form; stops at & or whitespace.
+_ACCESS_TOKEN_QS_RE = re.compile(
+    r"(access[_]?token\s*=\s*)[^&\s\"']+",
+    flags=re.IGNORECASE,
+)
+# Matches "access[_]?Token":"<value>" in JSON; stops at closing quote.
+_ACCESS_TOKEN_JSON_RE = re.compile(
+    r"(\"access[_]?token\"\s*:\s*\")[^\"]*",
+    flags=re.IGNORECASE,
+)
+
+
+def redact_text(text: str) -> str:
+    """Scrub bearer tokens and accessToken values from arbitrary text.
+
+    Pure function; no side effects. Used by `redact_error` and available
+    for use on any raw response body / stderr string before it reaches
+    a log or exception message.
+
+    renamed from `_redact_text` to a public symbol so
+    cross-module callers (sf_cli._redact_subprocess_stderr) import a stable
+    name instead of reaching into a module-private. The `_redact_text`
+    alias below is retained for backwards compatibility with any caller
+    that still imports the underscore-prefixed form; it will be removed
+    in a future batch once all callers migrate.
+    """
+    if not text:
+        return text
+    text = _AUTH_HEADER_RE.sub(r"\1<redacted>", text)
+    text = _ACCESS_TOKEN_QS_RE.sub(r"\1<redacted>", text)
+    text = _ACCESS_TOKEN_JSON_RE.sub(r'\1<redacted>', text)
+    return text
+
+
+# deprecated alias. Retained so existing tests and any lingering
+# `from rest_client import _redact_text` imports keep working. New code
+# MUST use `redact_text` (public). Planned removal in a follow-up batch
+# once the codebase is clean.
+_redact_text = redact_text
+
+
+def redact_error(exc: BaseException) -> str:
+    """Return a safe string representation of `exc`.
+
+    guaranteed to:
+      * include the exception type name (for triage)
+      * include the exception message WITH bearer tokens scrubbed
+      * NEVER include raw header collections, even if the exception carries them
+
+    For HTTPError specifically, we do NOT call `exc.read()` — the caller is
+    responsible for reading the body once (reading twice is usually a no-op
+    but we avoid the extra I/O and any token embedded in the body is scrubbed
+    at the caller via `redact_text` when it surfaces in the error message).
+    """
+    exc_type = type(exc).__name__
+    try:
+        raw = str(exc)
+    except Exception:
+        raw = "<unreprable>"
+    return f"{exc_type}: {redact_text(raw)}"
+
+
+# -----------------------------------------------------------------------------
+# cross-host redirect strips Authorization
+# -----------------------------------------------------------------------------
+
+
+class StripAuthOnCrossHostRedirect(urllib.request.HTTPRedirectHandler):
+    """Strip Authorization header on any cross-host redirect.
+
+    Python's default HTTPRedirectHandler preserves request headers across
+    301/302/303/307/308. When `instanceUrl` is the trusted origin and a
+    redirect points at ANY other hostname, we treat the Authorization
+    header as tainted and drop it before the follow-up request goes out.
+
+    Hostname comparison is case-insensitive (DNS is case-insensitive).
+    We compare `urlparse(...).hostname` — which strips port and userinfo
+    — because a port change on the same host is NOT a credential-leak
+    vector, but treating it as cross-host would break legitimate
+    `:443` → bare-host redirects.
+    """
+
+    def redirect_request(self, req, fp, code, msg, headers, newurl):
+        orig_host = urllib.parse.urlparse(req.full_url).hostname
+        new_host = urllib.parse.urlparse(newurl).hostname
+        # Normalize for case-insensitive comparison. `None == None` is safe
+        # (both malformed URLs treated as "same host" — urllib would fail
+        # the follow-up anyway).
+        same_host = (orig_host or "").lower() == (new_host or "").lower()
+
+        if same_host:
+            # Default behavior: preserve Authorization (same-host redirect is
+            # standard practice; stripping would break legitimate OAuth flows).
+            return super().redirect_request(req, fp, code, msg, headers, newurl)
+
+        # cross-host redirect — build a NEW Request without any
+        # Authorization header. We cannot mutate `req` in place because
+        # urllib's default handler reuses it; callers that retain a reference
+        # would observe the mutation.
+        new_req = super().redirect_request(req, fp, code, msg, headers, newurl)
+        if new_req is None:
+            return None
+        # Authorization may have been set via add_header (lowercased internal
+        # storage as "Authorization") or via unredirected_hdrs. Remove both.
+        # Note: `Request.headers` stores header names in the form produced by
+        # `.capitalize()` (so "Authorization" lands as "Authorization" since
+        # .capitalize() leaves single-word strings unchanged at the first
+        # letter — but we defensively pop both casings).
+        for hdr_name in list(new_req.headers.keys()):
+            if hdr_name.lower() == "authorization":
+                new_req.headers.pop(hdr_name, None)
+        for hdr_name in list(new_req.unredirected_hdrs.keys()):
+            if hdr_name.lower() == "authorization":
+                new_req.unredirected_hdrs.pop(hdr_name, None)
+        return new_req
+
+
+def build_opener() -> urllib.request.OpenerDirector:
+    """Return an OpenerDirector wired with StripAuthOnCrossHostRedirect.
+
+    every REST call in this module must go through an opener built
+    here. Direct use of `urllib.request.urlopen(...)` bypasses the redirect
+    handler and reintroduces the cross-host-token-leak vulnerability.
+    """
+    return urllib.request.build_opener(StripAuthOnCrossHostRedirect())
+
+
+# -----------------------------------------------------------------------------
+# 401 token refresh decorator + tooling/data query helpers
+# -----------------------------------------------------------------------------
+# Motivation: `sf org display` returns an access token with a short TTL
+# (typically 15min–2h). A pipeline run with many Flow / Tooling fetches can
+# exceed TTL. Without a refresh path the pipeline fails mid-run with a raw
+# 401 and the user has to re-run from scratch.
+#
+# Design: `retry_on_401` is a decorator-factory that takes a `refresh_fn`
+# closure. On 401 (HTTP status OR body contains INVALID_SESSION_ID) it
+# invokes `refresh_fn()`, retries the wrapped call ONCE, and surfaces the
+# original error if the retry also 401s. The refresh closure is passed
+# per-call rather than stored globally so the client stays stateless and
+# testable — each caller wires its own `lambda: run_sf("org_display", ...)`.
+#
+# Redaction: any error text that surfaces from this decorator runs through
+# `redact_error` / `redact_text` . The 401 body is read once and
+# scrubbed before being used for INVALID_SESSION_ID detection — we never
+# log or re-raise raw bytes from the wire.
+
+_INVALID_SESSION_MARKER = "INVALID_SESSION_ID"
+
+
+_ERROR_ENVELOPE_KEYS = ("errors", "error", "errorDetails", "messages")
+
+
+def _body_indicates_invalid_session(body: Any) -> bool:
+    """detect `INVALID_SESSION_ID` in a parsed or raw response body.
+
+    Some SF endpoints (notably a few Tooling paths) return HTTP 200 with an
+    error body rather than 401. The documented SF error shape is a list of
+    `{"errorCode": "INVALID_SESSION_ID", "message": "..."}` dicts at the
+    body root.
+
+    BUGFIX (2026-05-03): the prior implementation recursed into ALL
+    `dict.values()` and fell back to substring-matching the stringified
+    form. Tooling Query responses for ApexClass include a `records` list
+    whose `Body` field is the raw Apex source — which can legitimately
+    contain the literal string `INVALID_SESSION_ID` (as an errorCode
+    constant referenced by catch/rethrow code, e.g. `XCSF_FlowFaultMessage`
+    and `SkillRulesMatchAction` in the real-org fixture). That tripped the
+    value-walk and the substring fallback, misclassifying a 200 OK success
+    response as an auth failure and forcing a spurious
+    `INVALID_SESSION_ID after refresh` envelope into Wave B's `_unresolved`.
+
+    The tightened rules:
+      1. Top-level list/tuple: the SF error envelope — recurse (one level
+         is enough in practice, but recursion is safe because list items
+         are error dicts, not data rows with Apex bodies).
+      2. Dict: a positive match requires `errorCode == INVALID_SESSION_ID`
+         at THIS level, OR recursion into a well-known error-envelope key
+         (`errors`, `error`, `errorDetails`, `messages`). Data keys like
+         `records`, `Body`, `attributes` are NEVER walked — that was the
+         bug. Typos / shape variants still surface through the
+         error-envelope keys, which is the real world observed-shape set.
+      3. Top-level raw str/bytes: an unparsed error body. A substring
+         match here is still safe because this helper is only called on
+         the TOP-LEVEL parsed response (never on a field inside a success
+         dict), so a raw-string body means SF literally returned text
+         rather than JSON — in which case substring match is the best
+         we can do.
+      4. Anything else (int, float, None, other types): not an error
+         shape. Return False.
+    """
+    if body is None:
+        return False
+    if isinstance(body, (list, tuple)):
+        return any(_body_indicates_invalid_session(item) for item in body)
+    if isinstance(body, dict):
+        code = body.get("errorCode") or body.get("error_code")
+        if isinstance(code, str) and _INVALID_SESSION_MARKER in code:
+            return True
+        # Only recurse into known error-envelope keys. Walking arbitrary
+        # values misclassifies legitimate success payloads whose data
+        # fields (e.g. ApexClass.Body) happen to contain the string
+        # `INVALID_SESSION_ID`.
+        for key in _ERROR_ENVELOPE_KEYS:
+            if key in body and _body_indicates_invalid_session(body[key]):
+                return True
+        return False
+    if isinstance(body, (str, bytes)):
+        text = body.decode("utf-8", errors="replace") if isinstance(body, bytes) else body
+        return _INVALID_SESSION_MARKER in text
+    return False
+
+
+class _InvalidSessionSignal(Exception):
+    """Internal-only: tunnel INVALID_SESSION_ID-in-200-body through retry_on_401.
+
+    Never escapes this module — the decorator catches it, refreshes, retries,
+    and on a second failure surfaces a RestClientError (redacted).
+    """
+
+
+def _is_invalid_session_403(exc: urllib.error.HTTPError) -> bool:
+    """detect HTTP 403 + `INVALID_SESSION_ID` in body.
+
+    Some SF endpoints respond to an expired session with `403 Forbidden`
+    (body: `{"errorCode": "INVALID_SESSION_ID"}`) instead of 401. We treat
+    that shape as auth-refresh-triggering just like 401. Any OTHER 403
+    (permission denied, IP restriction, etc.) propagates unchanged — we
+    don't burn a refresh on a non-auth 403.
+
+    Reading `.read()` on an HTTPError is a one-shot — callers that surface
+    this exception must not re-read the body. For our retry path this is
+    fine: the decorator consumes the body once to decide, then discards
+    the exception after retry.
+    """
+    if exc.code != 403:
+        return False
+    try:
+        body = exc.read()
+    except Exception:
+        return False
+    if not body:
+        return False
+    try:
+        text = body.decode("utf-8", errors="replace") if isinstance(body, bytes) else str(body)
+    except Exception:
+        return False
+    return _INVALID_SESSION_MARKER in text
+
+
+def retry_on_401(
+    refresh_fn: Callable[[], Tuple[str, str]],
+) -> Callable[[Callable[..., Any]], Callable[..., Any]]:
+    """Decorator: retry a REST call ONCE after refreshing the session token.
+
+    detects three auth-failure shapes:
+      1. `urllib.error.HTTPError` with `code == 401`
+      2. `urllib.error.HTTPError` with `code == 403` AND body contains
+         `INVALID_SESSION_ID` (some SF endpoints return 403 for
+         expired sessions; non-auth 403 is NOT treated as refreshable).
+      3. HTTP 200 with response body containing `INVALID_SESSION_ID` —
+         callers signal this by raising `_InvalidSessionSignal`; the
+         helpers below do the body inspection before returning.
+
+     (REMEDIATE): the retry contract is documented and enforced here:
+    `refresh_fn()` returns `(instance_url, access_token)` and the caller
+    MUST thread those fresh credentials into the NEXT invocation of the
+    wrapped callable. See `tooling_query` / `data_query`, which implement
+    this via the `creds_provider` closure pattern (Option A). The previous
+    design captured credentials in a closure at call time — refresh
+    returned fresh creds, the wrapped closure kept using stale ones, and
+    the retry 401'd again. That's the defect. `retry_on_401` itself
+    is still credential-agnostic; the invariant lives in the caller.
+
+    If the retry ALSO fails (401 / 403+INVALID_SESSION_ID / INVALID_SESSION_ID),
+    the ORIGINAL error is re-raised (wrapped as `RestClientError` with a
+    redacted message) — not the retry's error. Rationale: the user sees
+    the first-attempt context (what call failed) rather than a downstream
+    artefact of the retry path.
+
+    Non-auth errors (500, 403 without INVALID_SESSION_ID, network timeout,
+    etc.) propagate unchanged — refresh is NOT attempted.
+    """
+
+    def _is_auth_http_error(exc: urllib.error.HTTPError) -> bool:
+        # 401 is the classic shape; 403+INVALID_SESSION_ID is a
+        # variant observed on some SF endpoints. All other 4xx/5xx
+        # propagate without refresh.
+        if exc.code == 401:
+            return True
+        if _is_invalid_session_403(exc):
+            return True
+        return False
+
+    def decorator(fn: Callable[..., Any]) -> Callable[..., Any]:
+        @functools.wraps(fn)
+        def wrapper(*args: Any, **kwargs: Any) -> Any:
+            try:
+                return fn(*args, **kwargs)
+            except urllib.error.HTTPError as first_exc:
+                if not _is_auth_http_error(first_exc):
+                    raise
+                # auth-failure path — refresh and retry once.
+                try:
+                    refresh_fn()
+                except Exception as refresh_exc:
+                    # Refresh itself failed — surface the REFRESH error so
+                    # the user knows they need to re-auth, but keep token
+                    # redaction on the message.
+                    raise RestClientError(
+                        f"token refresh failed after auth error: {redact_error(refresh_exc)}"
+                    ) from None
+                try:
+                    return fn(*args, **kwargs)
+                except urllib.error.HTTPError as retry_exc:
+                    if _is_auth_http_error(retry_exc):
+                        raise RestClientError(
+                            f"auth error after refresh (original): {redact_error(first_exc)}"
+                        ) from None
+                    raise
+                except _InvalidSessionSignal:
+                    raise RestClientError(
+                        f"INVALID_SESSION_ID after refresh (original): "
+                        f"{redact_error(first_exc)}"
+                    ) from None
+            except _InvalidSessionSignal as first_sig:
+                # body-path INVALID_SESSION_ID — refresh and retry once.
+                try:
+                    refresh_fn()
+                except Exception as refresh_exc:
+                    raise RestClientError(
+                        f"token refresh failed after INVALID_SESSION_ID: "
+                        f"{redact_error(refresh_exc)}"
+                    ) from None
+                try:
+                    return fn(*args, **kwargs)
+                except urllib.error.HTTPError as retry_exc:
+                    if _is_auth_http_error(retry_exc):
+                        raise RestClientError(
+                            f"auth error after refresh (original INVALID_SESSION_ID): "
+                            f"{redact_error(first_sig)}"
+                        ) from None
+                    raise
+                except _InvalidSessionSignal:
+                    raise RestClientError(
+                        f"INVALID_SESSION_ID after refresh: {redact_error(first_sig)}"
+                    ) from None
+
+        return wrapper
+
+    return decorator
+
+
+# -----------------------------------------------------------------------------
+# transient HTTP (429/503) exponential backoff
+# -----------------------------------------------------------------------------
+# Motivation: SF REST endpoints occasionally return 429 (rate-limited) or
+# 503 (service unavailable) during API-limit windows, planned maintenance,
+# or transient edge hiccups. Without a retry the pipeline fails mid-run on
+# what is almost always a recoverable condition. Three attempts with
+# exponential backoff (1s → 2s → 4s, roughly — `base_delay * 2**attempt`)
+# resolve the vast majority of transient blips without meaningfully
+# extending happy-path latency.
+#
+# Scope (what this decorator does NOT do):
+# * Not a 401 refresher — that's `retry_on_401`. The layering is:
+# retry_on_401( retry_on_transient_http()( _query_once ) )
+# so 429s retry first (inner), and a 401 surfacing through that layer
+# triggers the outer refresh. The two concerns stay independent.
+# * No jitter. Bounded to 3 attempts + short delays; adding jitter here
+# would only matter if many callers burst-retried against the same
+# endpoint, which is not the shape of this skill's traffic.
+# * No retry on 5xx generally. Only 503 is retried because 500/502/504
+# often indicate deterministic server-side failures where retrying
+# won't help and we'd prefer to surface fast.
+#
+# Retry-After handling:
+# * If the header is present and parses as a non-negative number of
+# seconds OR an HTTP-date, we honor it — bounded below by base_delay *
+# 2**attempt so a "Retry-After: 0" doesn't disarm the backoff.
+# * Negative / malformed values fall back to the exponential schedule.
+#
+# Redaction: any HTTPError surfaced from the final failure goes through
+# `redact_error` at the caller's message site (see `_query_once`'s wrapper
+# chain). We do NOT stringify headers here — no `str(exc.headers)` or
+# `resp.read()` debug dumps on retry. URL logging strips the query string.
+
+
+def _parse_retry_after(raw: str | None) -> float | None:
+    """Parse a Retry-After header value to seconds.
+
+    Per RFC 9110 the value is either `delta-seconds` (non-negative int) or
+    an HTTP-date. We accept floats too — some clients emit decimal seconds
+    and the spec-strict int parse would silently lose them.
+
+    Returns None for missing / malformed input; caller falls back to the
+    exponential schedule.
+    """
+    if not raw:
+        return None
+    raw = raw.strip()
+    try:
+        secs = float(raw)
+        if secs < 0:
+            return None
+        return secs
+    except ValueError:
+        pass
+    # HTTP-date form. `parsedate_to_datetime` raises on malformed input on
+    # 3.10+; wrap defensively.
+    try:
+        target = email.utils.parsedate_to_datetime(raw)
+    except (TypeError, ValueError):
+        return None
+    if target is None:
+        return None
+    # Header expresses an absolute time; convert to a delta from now.
+    now = time.time()
+    delta = target.timestamp() - now
+    if delta < 0:
+        return 0.0
+    return delta
+
+
+def retry_on_transient_http(
+    max_retries: int = 3,
+    base_delay: float = 1.0,
+) -> Callable[[Callable[..., Any]], Callable[..., Any]]:
+    """Decorator: retry on HTTP 429/503 up to `max_retries` times.
+
+    on `urllib.error.HTTPError` with code in {429, 503}, sleep and
+    retry. Any other exception (including 401 — let `retry_on_401`
+    handle that, and any 5xx besides 503) propagates immediately.
+
+    Retry semantics:
+      * max_retries: number of retries AFTER the first attempt (default 3).
+      * Total attempts = 1 + max_retries (default 4: 1 original + 3 retries).
+      * Final attempt's exception propagates; no further retry.
+
+    Delay per attempt = max(Retry-After-seconds, base_delay * 2**attempt).
+    If Retry-After is absent or malformed, falls back to the exponential
+    schedule. `attempt` is 0-indexed for the first retry (so the first
+    sleep is base_delay * 1 = base_delay).
+
+    Logs a DEBUG breadcrumb per retry via the module-level `logger`. No
+    header values are logged — only the HTTP code + computed delay +
+    attempt counter.
+    """
+
+    def decorator(fn: Callable[..., Any]) -> Callable[..., Any]:
+        @functools.wraps(fn)
+        def wrapper(*args: Any, **kwargs: Any) -> Any:
+            # Total attempts = 1 original + max_retries retries. `attempt`
+            # below is 0-indexed over the retry slots, which matches the
+            # exponential formula base_delay * 2**attempt starting at
+            # base_delay for attempt=0.
+            for attempt in range(max_retries):
+                try:
+                    return fn(*args, **kwargs)
+                except urllib.error.HTTPError as exc:
+                    if exc.code not in (429, 503):
+                        # Any other HTTP status (including 401) — out of
+                        # scope for this decorator. Propagate so outer
+                        # layers (retry_on_401) can handle it.
+                        raise
+                    retry_after_raw = None
+                    try:
+                        # `exc.headers` is a Message-like; .get tolerates
+                        # the case where headers are absent on some
+                        # synthesized HTTPErrors.
+                        retry_after_raw = exc.headers.get("Retry-After") if exc.headers else None
+                    except Exception:
+                        retry_after_raw = None
+                    hinted = _parse_retry_after(retry_after_raw)
+                    expo = base_delay * (2 ** attempt)
+                    delay = max(hinted, expo) if hinted is not None else expo
+                    logger.debug(
+                        "retry_on_transient_http: HTTP %d, sleeping %.2fs (attempt %d/%d)",
+                        exc.code, delay, attempt + 1, max_retries,
+                    )
+                    time.sleep(delay)
+                    continue
+            # Final attempt — let any exception propagate directly. If
+            # this attempt raises 429/503 the caller gets the HTTPError
+            # unchanged (redaction happens at the wrapping call site).
+            return fn(*args, **kwargs)
+
+        return wrapper
+
+    return decorator
+
+
+def _query_once(
+    instance_url: str,
+    token: str,
+    path: str,
+    soql: str,
+) -> dict:
+    """Single SOQL GET — reads body, checks for INVALID_SESSION_ID.
+
+    the Salesforce REST Query + Tooling Query
+    endpoints are GET-only with the SOQL passed as a urlencoded `q=`
+    querystring. The prior POST-with-JSON-body shape returned HTTP 405
+    ("Method Not Allowed") on every real-org run — confirmed by curl
+    against a real org:
+
+        GET  /services/data/v60.0/query/?q=<soql>  -> 200 OK
+        POST /services/data/v60.0/query/           -> 405 Method Not Allowed
+
+    We urlencode the SOQL (via `urllib.parse.urlencode({"q": soql})`)
+    rather than string-concatenating so spaces, single quotes, commas,
+    and parentheses in the query — all legal inside SOQL string
+    literals — travel through the wire correctly. The JSON body and
+    `Content-Type: application/json` header are dropped (GET carries
+    no body; Content-Type with a bodyless GET is technically legal but
+    meaningless and some edge proxies get irritated by it).
+
+    uses `build_opener()` so cross-host redirects strip Authorization.
+    any exception is surfaced with `redact_error`; raw body is run
+    through `redact_text` before it reaches an exception message.
+    signals `_InvalidSessionSignal` on 200-body-error so
+    `retry_on_401` can refresh + retry.
+    sets `Accept-Encoding: identity` to opt out of gzip. The
+    responses here are small JSON payloads — gzip would add handling
+    complexity (Content-Encoding detection + decompression) for zero
+    measurable benefit. Explicit `identity` keeps the body-inspection
+    path (INVALID_SESSION_ID detection) straightforward and testable.
+    """
+    # urlencode handles all SOQL-legal characters — spaces,
+    # single quotes inside string literals, commas, parentheses. Hand-
+    # rolled concatenation would have to escape each class separately.
+    qs = urllib.parse.urlencode({"q": soql})
+    url = f"{instance_url.rstrip('/')}{path}?{qs}"
+    req = urllib.request.Request(url, method="GET")
+    req.add_header("Authorization", f"Bearer {token}")
+    req.add_header("Accept", "application/json")
+    # force identity encoding — no gzip.
+    req.add_header("Accept-Encoding", "identity")
+
+    opener = build_opener()
+    try:
+        with opener.open(req) as resp:
+            raw = resp.read()
+    except urllib.error.HTTPError as exc:
+        # Bug D.1 fix: HTTPError.read() is one-shot. The default
+        # `redact_error()` path stringifies just `HTTP Error 4xx: <reason>`
+        # which loses the Salesforce error body — and that body is what
+        # tells the operator WHY (e.g. INVALID_FIELD, MALFORMED_QUERY,
+        # name-too-long, or a specific bad identifier). Pull the body
+        # ONCE here, scrub it via redact_text, attach it to the exception
+        # as `_response_body_preview`, and re-raise. Downstream callers
+        # that surface `redact_error(exc)` keep their existing string
+        # output; callers that want the body read it from the attribute.
+        try:
+            body = exc.read()
+        except Exception:
+            body = b""
+        try:
+            body_text = body.decode("utf-8", errors="replace")
+        except Exception:
+            body_text = ""
+        # Cap at 500 chars — enough to carry a Salesforce error array
+        # element with the offending name in it; small enough that
+        # downstream contexts don't bloat. redact_text strips bearer
+        # tokens defensively (body shouldn't contain one but cheaper to
+        # always scrub than reason about it).
+        exc._response_body_preview = redact_text(body_text)[:500]  # type: ignore[attr-defined]
+        raise
+
+    try:
+        parsed = json.loads(raw.decode("utf-8"))
+    except (UnicodeDecodeError, json.JSONDecodeError) as e:
+        raise RestClientError(
+            f"malformed query response: {redact_error(e)}"
+        ) from None
+
+    # HTTP 200 + INVALID_SESSION_ID in body → tunnel up to the
+    # decorator so it can refresh + retry.
+    if _body_indicates_invalid_session(parsed):
+        raise _InvalidSessionSignal(
+            redact_text(json.dumps(parsed)[:500])
+        )
+    return parsed
+
+
+# (REMEDIATE): credentials flow through a provider closure, not
+# through function arguments that are captured once at decoration time.
+#
+# The previous design had a fatal defect:
+#
+# def tooling_query(instance_url, token, soql, *, on_401_refresh):
+# @retry_on_401(on_401_refresh)
+# def _call():
+# return _query_once(instance_url, token, ...) # stale closure
+# return _call()
+#
+# `refresh_fn()` returned `(new_url, new_token)` but those values were
+# thrown away — the decorator only called `refresh_fn()` for its side
+# effects, then re-invoked the same closure holding the ORIGINAL stale
+# `instance_url` / `token`. Retry 401'd again. Feature was dead.
+#
+# Option A (taken): `tooling_query` / `data_query` accept a
+# `creds_provider: Callable[[], Tuple[str, str]]` that is called EACH
+# attempt. `refresh_fn` is responsible for mutating whatever state the
+# provider reads from — typically a simple closure over a list. This
+# keeps `retry_on_401` credential-agnostic and makes the contract
+# ("refresh updates the source that `creds_provider` reads from")
+# explicit at the caller site instead of implicit in the helper.
+#
+# Test: `RetryOn401CredentialRefreshIntegrationTests` in
+# test_rest_client.py verifies that a real refresh carries fresh
+# credentials into the retry call; the prior design fails that test.
+
+
+def tooling_query(
+    creds_provider: Callable[[], Tuple[str, str]],
+    soql: str,
+    *,
+    api_version: str,
+    on_401_refresh: Callable[[], Tuple[str, str]],
+) -> dict:
+    """GET a SOQL query against the Tooling API.
+
+    `creds_provider()` is invoked on EACH attempt. `on_401_refresh`
+    is responsible for updating whatever state the provider reads from
+    before it returns — otherwise the retry would hit the same stale
+    credentials and 401 again.
+
+    `api_version` is a REQUIRED keyword-only arg.
+    The prior hardcoded `v60.0` was the source of — real orgs
+    run on v66 and expose fields v60 does not (confirmed empirically:
+    `BotDefinition.Description` exists on v66, `INVALID_FIELD` on v60).
+    Callers thread the version through from `main._derive_org_ids`, which
+    reads it once from `sf org display --json`. Making the param REQUIRED
+    (no default) surfaces a missed call-site as a TypeError at call time
+    rather than a silent regression back to a stale pinned version.
+    Shape is already enforced by `fs_guard.validate_api_version`
+    (`^v[0-9]+\\.[0-9]+$`) at the caller.
+
+    Stateless: both closures are passed per-call. Each caller owns the
+    storage the refresh writes to (typically a small `list[tuple[str, str]]`
+    cell).
+    """
+    path = f"/services/data/{api_version}/tooling/query/"
+
+    # retry_on_transient_http is the INNER layer — 429/503 retries
+    # happen first, and any 401 bubbling through triggers retry_on_401's
+    # refresh path at the outer layer. Ordering matters: if transient
+    # retry were outer, a 429 during a refresh retry would be mis-handled.
+    @retry_on_401(on_401_refresh)
+    @retry_on_transient_http()
+    def _call() -> dict:
+        # re-read creds on every attempt so refresh actually lands.
+        instance_url, token = creds_provider()
+        return _query_once(instance_url, token, path, soql)
+
+    return _call()
+
+
+def data_query(
+    creds_provider: Callable[[], Tuple[str, str]],
+    soql: str,
+    *,
+    api_version: str,
+    on_401_refresh: Callable[[], Tuple[str, str]],
+) -> dict:
+    """GET a SOQL query against the Data API (non-Tooling).
+
+    identical retry wiring to `tooling_query`, different URL path.
+    `creds_provider` is invoked on every attempt so a refresh actually
+    propagates fresh credentials into the retry.
+
+    `api_version` is a REQUIRED keyword-only arg;
+    see `tooling_query` for the rationale.
+    """
+    path = f"/services/data/{api_version}/query/"
+
+    # see `tooling_query` for the decorator-ordering rationale.
+    @retry_on_401(on_401_refresh)
+    @retry_on_transient_http()
+    def _call() -> dict:
+        # re-read creds on every attempt so refresh actually lands.
+        instance_url, token = creds_provider()
+        return _query_once(instance_url, token, path, soql)
+
+    return _call()
+
+
+def static_creds(instance_url: str, token: str) -> Callable[[], Tuple[str, str]]:
+    """Build a zero-state creds_provider that always returns the same pair.
+
+    Convenience for callers that don't wire a refresh path — tests,
+    single-shot scripts where a 401 is terminal. If the call 401s, the
+    retry sees the same creds and fails again; this is correct behavior
+    for an unrefreshable context (no point claiming otherwise).
+    """
+    def _provider() -> Tuple[str, str]:
+        return instance_url, token
+    return _provider