framework: port R4 credential-helper pair from brain 15f6979a (python entrypoint w/ alpine bash-path candidates + impl, byte-faithful; P0-SEC T125)

(cherry picked from commit 13e7398c2a)
This commit is contained in:
code-be-01
2026-08-30 16:51:40 -05:00
committed by marcie
parent f194ccb8a3
commit a47f7aec10
2 changed files with 536 additions and 212 deletions
@@ -1,220 +1,64 @@
#!/bin/bash
# git-credential-mosaic — git credential helper. Resolves a Gitea token from the
# Mosaic credential store at runtime so remote URLs never embed secrets.
#!/usr/bin/python3
# git-credential-mosaic — production entrypoint (P0-SEC R4, rev-code-02 B1).
#
# Install (one-time, per clone or globally):
# git config credential.helper "$HOME/.config/mosaic/tools/git/git-credential-mosaic"
# WHY THIS IS NOT BASH: three review rounds falsified every in-bash startup
# guard. A non-interactive bash sources $BASH_ENV and imports exported
# functions BEFORE the first script line, so read(), unset(), exit(),
# declare(), printf() — every callable — can be shadows that fake the
# ancestry, defeat the scrub, or forge diagnostics (rev-code-02 probes 1 and
# 2, artifacts fc49e9d9 lineage). No in-language dispatch survives that.
#
# Per-agent identity (Gate-16 author != reviewer separation):
# git config mosaic.gitIdentity <agent-id> # per-worktree, persists on disk
# # or: export MOSAIC_GIT_IDENTITY=<agent-id>
# This entrypoint is unshapable at the bash level: python does not read
# BASH_ENV and imports no bash functions, and the interpreter is pinned by
# absolute shebang (no PATH resolution). It builds the child environment BY
# ALLOWLIST and execve's the bash implementation directly — the child bash
# starts with no BASH_ENV, no BASH_FUNC_*, no SHELLOPTS/BASHOPTS, and exactly
# the variables the credential protocol needs. stdin/stdout/stderr and argv
# pass through untouched.
#
# ── WHY THIS FAILS CLOSED ──────────────────────────────────────────────────────
# This helper used to end by emitting the shared account's token for any request
# it could not resolve to an identity. A seat with no identity, or with an
# identity whose token was never provisioned, therefore received the most
# privileged credential configured on the host — silently, and indistinguishably
# from correct operation. Every record it then created (commit, push, PR, review)
# was attributed to that shared account, so author != reviewer separation was
# unenforceable and the true actor was unrecoverable after the fact.
#
# Under-provisioning must fail loudly, not impersonate. A refused git operation
# is recoverable in one command; a merged pull request attributed to the wrong
# principal is not.
#
# ── CONTRACT ───────────────────────────────────────────────────────────────────
# identity : MOSAIC_GIT_IDENTITY > git config mosaic.gitIdentity > the
# username git supplies on stdin
# store : chosen by what the identity IS, with no precedence and no
# cross-store fallback (see "Credential store selection" below)
# hit : emit username + password, exit 0
# miss : emit NOTHING, spool a durable escalation record, explain on
# stderr, exit 1 — git surfaces the failure and nothing is attributed
# unknown host : exit 0 with no output, no record (passthrough for non-Mosaic
# remotes handled by another helper)
#
# Backward compatibility is preserved for exactly one case: a host with no fleet
# and no identity requested still gets the shared account, because on such a host
# the shared account is the operator's own and there is no attribution to lose.
# A host that HAS a fleet has agents whose records must be distinguishable, so
# the shared fallback is refused there.
#
# A token is never written to stderr, to the escalation record, or to any log.
# The implementation file (git-credential-mosaic.impl) refuses to run without
# the clean-mode marker, so it cannot be invoked directly as a shaped-entry
# bypass of this wrapper.
[ "$1" = "get" ] || exit 0
import os
import sys
host=""; username_in=""
while IFS= read -r line; do
[ -z "$line" ] && break
case "$line" in
host=*) host=${line#host=};;
username=*) username_in=${line#username=};;
esac
done
IMPL = os.path.join(os.path.dirname(os.path.realpath(__file__)), "git-credential-mosaic.impl")
# Absolute-path candidates ONLY — never PATH resolution (an attacker-shaped
# PATH must not choose the interpreter). /usr/bin/bash is the fleet-host
# layout; /bin/bash is alpine and other FHS variants (found by the T125
# gateway-image verification: the hardcoded /usr/bin/bash made every call
# exit 127 inside node:22-alpine).
BASH_CANDIDATES = ("/usr/bin/bash", "/bin/bash")
BASH = next((p for p in BASH_CANDIDATES if os.access(p, os.X_OK)), None)
# Recognized Gitea hosts carry the per-identity token scheme. Anything else is
# declined quietly — another helper owns it, and refusing would break it.
case "$host" in
git.uscllc.com) idpfx=gitea-usc;;
git.mosaicstack.dev) idpfx=gitea-mosaicstack;;
*) exit 0;;
esac
# Allowlist: everything else in the environment dies at this boundary. Adding
# a variable here is a security decision — it crosses into a shell that no
# longer has any startup shaping, but it also becomes the only context the
# implementation can see.
KEEP = (
"HOME",
"PATH",
"LANG",
"MOSAIC_GIT_IDENTITY",
"MOSAIC_AGENT_NAME",
"MOSAIC_BRAIN_HOME",
"MOSAIC_CREDENTIAL_SPOOL",
"MOSAIC_CREDENTIAL_LINEAGE_FENCE",
)
ident="$MOSAIC_GIT_IDENTITY"; ident_src="MOSAIC_GIT_IDENTITY"
if [ -z "$ident" ]; then
ident=$(git config --get mosaic.gitIdentity 2>/dev/null)
ident_src="git config mosaic.gitIdentity"
fi
if [ -z "$ident" ]; then
ident="$username_in"
ident_src="the username git supplied"
fi
env = {"_MOSAIC_HELPER_CLEAN": "1"}
for name in KEEP:
value = os.environ.get(name)
if value is not None:
env[name] = value
# ── Credential store selection ────────────────────────────────────────────────
# An identity is a SEAT or it is a SERVICE, and which one it is determines where
# its credential lives. There is no precedence rule between the two stores and no
# fallback from one to the other: a seat whose slot is empty fails closed rather
# than reading a service credential that happens to share its name.
#
# seat — <brain>/fleet/agents/<ident>/ exists
# credential at <brain>/fleet/agents/<ident>/secrets/<idpfx>-<ident>.token
# service — it does not
# credential at ~/.config/mosaic/secrets/gitea-tokens/<idpfx>-<ident>.token
#
# One credential, one location. Two copies of one credential diverge, and the
# stale copy fails in a way that reads as a revoked token rather than as drift.
#
# Brain-home resolution mirrors packages/mosaic/src/fleet/brain-home.ts and
# tools/fleet/start-agent-session.sh: MOSAIC_BRAIN_HOME wins, else ~/.mosaic.
brain_home="${MOSAIC_BRAIN_HOME:-$HOME/.mosaic}"
svc_store="$HOME/.config/mosaic/secrets/gitea-tokens"
idtok=""; ident_kind=""
if [ -n "$ident" ]; then
if [ -d "$brain_home/fleet/agents/$ident" ]; then
ident_kind="seat"
idtok="$brain_home/fleet/agents/$ident/secrets/${idpfx}-${ident}.token"
else
ident_kind="service identity"
idtok="$svc_store/${idpfx}-${ident}.token"
fi
if [ -r "$idtok" ]; then
echo "username=${ident}"
echo "password=$(cat "$idtok")"
exit 0
fi
fi
# ── Shared-account fallback: ONLY on a host with no fleet and no identity ──────
# `fleet/agents` existing is the same signal brain-home.ts uses to decide a brain
# is active. Where there are seats, records must be attributable, so an
# unresolvable request is refused instead of borrowing the shared account.
fleet_present=0
[ -d "$brain_home/fleet/agents" ] && fleet_present=1
if [ -z "$ident" ] && [ "$fleet_present" -eq 0 ]; then
script_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# shellcheck source=../_lib/credentials.sh
source "$script_dir/../_lib/credentials.sh"
load_credentials "$idpfx" >/dev/null 2>&1 || exit 0
# GITEA_USER is not populated by load_credentials (it exports GITEA_URL and
# GITEA_TOKEN only). Gitea's git-over-HTTP auth authenticates from the token in
# the password field, not from the username string, so any non-empty
# placeholder works — deliberately NOT a real account name, since framework
# files stay operator-agnostic (tools/quality/scripts/verify-sanitized.sh).
echo "username=${GITEA_USER:-git}"
echo "password=$GITEA_TOKEN"
exit 0
fi
# ── FAIL CLOSED ───────────────────────────────────────────────────────────────
if [ -z "$ident" ]; then
reason="no-identity"
else
reason="no-token-for-identity"
fi
seat="${MOSAIC_AGENT_NAME:-unknown}"
ts=$(date -u +%Y-%m-%dT%H:%M:%SZ)
# The escalation RECORD is durable and unconditional; any notification built on
# top of it is best-effort. Record and alert are deduplicated separately — a cap
# on the alert alone lets the spool grow without bound exactly while the operator
# is being told nothing, so the louder the failure the quieter it gets.
#
# A record field is arbitrary operator-supplied text: an identity comes from git
# config or the environment, and cwd is whatever directory git ran in. Either can
# contain a quote or a backslash, which would make the line unparseable JSON --
# and a spool that silently stops parsing is worse than no spool, because the
# operator only discovers it while reading the record that explains an outage.
json_escape() {
local s=$1
s=${s//\\/\\\\}
s=${s//\"/\\\"}
s=${s//$'\t'/\\t}
s=${s//$'\r'/\\r}
s=${s//$'\n'/\\n}
printf '%s' "$s"
}
spool="${MOSAIC_CREDENTIAL_SPOOL:-$HOME/.local/state/mosaic-credential-escalations}"
spool_record=""
if mkdir -p "$spool" 2>/dev/null; then
chmod 700 "$spool" 2>/dev/null
spoolfile="$spool/$(date -u +%Y%m%d).jsonl"
dedupe="$spool/.spooled-${seat}-${ident:-none}-${reason}-$(date -u +%Y%m%d%H%M)"
if [ ! -e "$dedupe" ]; then
: > "$dedupe" 2>/dev/null
printf '{"ts":"%s","reason":"%s","identity":"%s","identity_source":"%s","kind":"%s","seat":"%s","host":"%s","cwd":"%s"}\n' \
"$(json_escape "$ts")" "$(json_escape "$reason")" \
"$(json_escape "${ident:-<unset>}")" "$(json_escape "$ident_src")" \
"$(json_escape "${ident_kind:-none}")" "$(json_escape "$seat")" \
"$(json_escape "$host")" "$(json_escape "$PWD")" \
>> "$spoolfile" 2>/dev/null
chmod 600 "$spoolfile" 2>/dev/null
fi
# Name the record only if one is actually on disk. Printing the path
# unconditionally sends the operator to a file that does not exist on exactly
# the hosts where the spool could not be created.
[ -s "$spoolfile" ] && spool_record="$spoolfile"
find "$spool" -maxdepth 1 -name '.spooled-*' -mmin +120 -delete 2>/dev/null
fi
cat >&2 <<EOF
git-credential-mosaic: REFUSED (fail-closed).
host : ${host}
identity : ${ident:-<unset>}${ident:+ (from ${ident_src}; resolved as a ${ident_kind})}
reason : ${reason}
EOF
if [ -n "$ident" ]; then
cat >&2 <<EOF
expected : ${idtok}
EOF
fi
cat >&2 <<EOF
No per-identity credential resolved. This helper does NOT fall back to the shared
account: that fallback makes every record it creates attributable to one
principal, which is unrecoverable once a pull request has merged under it.
Fix (pick one):
export MOSAIC_GIT_IDENTITY=<agent-id> # process-scoped
git config mosaic.gitIdentity <agent-id> # per-repo/worktree, persists
Then provision that identity's credential at the path named above. An identity
with a directory under \${MOSAIC_BRAIN_HOME:-\$HOME/.mosaic}/fleet/agents/ is a
seat and is read ONLY from its own secrets/ slot; any other identity is read from
~/.config/mosaic/secrets/gitea-tokens/. There is no fallback between the two.
If this identity legitimately needs git access and has none, ask the orchestrator
to provision one.
EOF
if [ -n "$spool_record" ]; then
echo " record: ${spool_record}" >&2
else
echo " record: NOT WRITTEN — spool unavailable at ${spool}" >&2
fi
exit 1
argv = [BASH, IMPL] + sys.argv[1:]
if BASH is None:
sys.stderr.write("git-credential-mosaic: no executable bash at " + " or ".join(BASH_CANDIDATES) + "\n")
sys.exit(127)
try:
os.execve(BASH, argv, env)
except OSError as exc:
sys.stderr.write(f"git-credential-mosaic: entrypoint exec failed: {exc}\n")
sys.exit(127)